The Invisible Burden: Examining the Impact of Exposure Misclassification in Epidemiologic Analyses of Uterine Fibroids.

J.D.D., N.R.Z.S., S.N.H., E.F.S., A.D., S.L.M., and E.C.C. contributed to the conception of the research question and study design. J.D.D. performed the analysis. All authors contributed to the interpretation of results and approved the final version of the manuscript.

In epidemiologic studies of uterine fibroids, investigators often rely on low‐sensitivity ascertainment methods such as self‐report. In this simulation study examining exposure misclassification of uterine fibroids, we demonstrate that the observed RR will be biased towards the null under reasonable scenarios. In cases of very low sensitivity, such as that for self‐reported status of uterine fibroids, a true effect estimate may be attenuated to the null, leading investigators to miss important, true effects. This could have deleterious impacts on the interpretation of study estimates, as investigators may conclude uterine fibroids have no effect on health outcomes even though one might exist. This simulation study comprehensively examined uterine fibroid exposure misclassification. We considered both non‐differential and differential misclassification, analysed variations based on fibroid symptom presence, and assessed sensitivity differences across racial and ethnic groups. Additionally, the robust design incorporated multiple variable parameters—including fibroid population prevalence, true relative risks and detection method sensitivity and specificity—enabling exploration of a diverse range of potential scenarios. This simulation study is limited in that we focused on largely simple scenarios, and our assumptions may not fully capture the true nature of data that investigators may encounter. For instance, despite the population prevalence of uterine fibroids being directly correlated with the age of females sampled, we did not consider the differential sensitivity of ascertainment methods by age. Nevertheless, we aimed to examine the impact of some of these factors (i.e., severity of uterine fibroids, true population prevalence) on the degree and direction of bias association with misclassification. Second, we did not account for how sampling decisions in studies utilising low‐ascertainment methods could introduce additional biases that may not be correctable through QBA. For example, studies relying on electronic health records often restrict their samples to individuals who underwent imaging, potentially introducing selection bias and overrepresenting other gynecologic conditions that warranted imaging. Such systemic biases fall outside the scope of correction through QBA, though are important to consider for fibroid research. Furthermore, some investigators may focus on fibroids as the outcome rather than the exposure. Investigators may also study fibroids as an outcome rather than an exposure. As approaches to addressing measurement error are similar for exposures and outcomes, ensuring accurate fibroid ascertainment during study design or conducting sensitivity analyses to account for measurement error remains crucial in such studies. Finally, although previous research has investigated uterine fibroids as a potential hypertension risk factor, the true relationship between these conditions remains uncertain and could reflect shared causal factors rather than direct causation. However, our primary aim in utilising this example was to provide a practical, real‐world application of our simulation for illustrative purposes. Chen et al. [ 27 ], who found a 1.35‐fold increased risk of hypertension, used electronic medical records to identify uterine fibroids in the cohort, whereas Haan et al. relied on self‐report and estimated a larger association of 1.80 [ 9 ]. Neither study used the gold standard for fibroid ascertainment or considered how the symptom status of uterine fibroids or population demographic characteristics might impact results. Of note, these two studies are not directly comparable due to different study population restrictions and adjustments for confounders. However, assuming confounding was adequately accounted for and that detection bias (where those with hypertension are more likely to have fibroids detected) did not distort a true null effect, we found that these observed effect estimates are likely underestimates of the true strength of the relationship. Importantly, these two studies employed cross‐sectional designs, although Haan et al. collected self‐reported fibroid history while measuring outcomes through blood pressure readings taken during the visit. However, establishing a temporal sequence is essential for causal inference, and even perfect variable classification within a cross‐sectional framework cannot establish causality. Differential misclassification by racialized groups was also examined. Racial disparities in the diagnosis of uterine fibroids are well described, though underlying mechanisms are unclear [ 1 , 3 , 6 ]. Black females tend to develop uterine fibroids at younger ages [ 18 , 31 , 32 ], which can lead to larger, more symptomatic fibroids, evidenced by the disproportionately high rate of hysterectomies for the removal of fibroids among Black females [ 13 , 14 ]. Such disparity in the severity of symptoms could increase sensitivity, particularly among subjective ascertainment mechanisms (e.g., self‐report), considering that many females with uterine fibroids, though symptomatic, live with severe symptoms and delay seeking care [ 13 , 14 ]. These disparities can introduce bias in epidemiologic studies. Specifically, when sensitivity was 50% or lower, and the overall population prevalence of uterine fibroids was high, the observed RR for Black females was more biased towards the null than for White females. Consequently, studies relying on self‐reported diagnoses of uterine fibroids may fail to identify a true effect among Black females while identifying it among White females, potentially worsening disparities. Studies using self‐report methods are subject to differential misclassification, which may lead to observed differences being attributable to varying diagnostic sensitivity rather than the truth. As uterine fibroids disproportionately impact Black females [ 33 , 34 ], this could worsen health disparities and result in missed opportunities for targeted intervention. However, if the positive finding stems from detection bias, increased intervention could lead to unnecessary surgical treatments without a true effect. These possibilities underscore the importance of critically assessing bias and using rigorous methods to address it, ensuring evidence‐based intervention. In the design phase of primary data collection, using high‐sensitivity ascertainment methods is critical to reduce bias. For example, the Study of Environment, Lifestyle & Fibroids (SELF) screened Black females aged 23–35 in Michigan with ultrasounds at enrollment and follow‐up to examine fibroid development [ 35 ]. However, such methods are costly and may not be feasible for large cohorts. Hybrid designs, combining subjective methods like self‐report or ICD codes with a subset undergoing gold‐standard diagnostics (e.g., transvaginal ultrasound), can correct exposure misclassification. When primary data collection is unavailable, studies rely on secondary analyses of cohorts using self‐reported fibroid status or ‘real‐world’ data, such as electronic health records, which often capture symptomatic cases requiring medical management. QBA can address these limitations, as shown in this simulation study. Sensitivity and specificity parameters (Table  1 ) can adjust observed estimates for measurement error and address race‐based misclassification. Increased use of QBA and sensitivity analyses in epidemiologic studies has been recommended [ 30 ]. Even so, the rigour of QBA is directly related to how closely the parameters reflect the truth. Although the truth may never be known, increasing the number of estimated parameters from validation studies will reduce uncertainty in the evidence. Validation studies on ascertainment methods of uterine fibroids, most specifically self‐report and ICD codes, are limited, and investment in future validation studies is needed to enhance robustness in the field of gynecologic epidemiology. These parameters can be used to correct for misclassification from biased exposure ascertainment. Although this simulation study centred on uterine fibroids, the use of QBA to adjust effect estimates for measurement error and the need for validation studies on ascertainment methods will improve the rigour of epidemiologic studies on other non‐malignant gynecologic conditions, such as polycystic ovary syndrome, endometriosis and adenomyosis.

Prevalence of uterine fibroids and validation of ascertainment methods: Estimates of uterine fibroid prevalence and incidence vary due to differences in assessment ages and methods (Table  1 ). Fibroid growth, influenced by oestrogen and progesterone, begins during reproductive years and often regresses by menopause [ 1 ]. Select examples of sensitivity and specificity of ascertainment methods for uterine fibroids. Abbreviations: ICD, International Classification of Disease; MRI, magnetic resonance imaging. The gold standard comparison to ICD codes used by Payne et al. was medical record review, which is not the gold standard for detection of uterine fibroids, and therefore these parameters are likely overestimates of the sensitivity and specificity of ICD codes. Sensitivity = Probability [Identifying uterine fibroids | Females with uterine fibroids]. Specificity = Probability [Identifying no uterine fibroids | Females without uterine fibroids]. US registry studies report incidence rates of uterine fibroids from 217 cases per 100,000 person‐years (diagnoses codes) to 3745 cases per 100,000 person‐years (hysterectomy patients) [ 1 , 22 , 23 ]. Prevalence ranges from 4.5% (self‐report among females aged 18–65) to 68.6% (transvaginal ultrasound among females 40–47 years old) [ 8 , 24 ]. Of note, validation studies often focus on larger fibroids or hysterectomy patients [ 20 , 21 ], underestimating true incidence and diagnostic sensitivity. Transvaginal ultrasounds and pelvic MRIs detect fibroids with 99% sensitivity [ 20 ], but are rarely used in epidemiology studies due to cost and invasiveness. Self‐reported methods, often based on pelvic exams suggesting an enlarged uterus, have a sensitivity lower than 50% [ 19 ]. Specificity is high (> 86%) among all methods [ 19 , 20 , 25 ], but misclassification remains a concern, particularly for asymptomatic cases, which comprise 75% of fibroid cases [ 19 , 26 ]. Relationship between uterine fibroids and hypertension: Two studies suggest a 1.3–1.8 increased risk of hypertension among females with fibroids compared to those without [ 9 , 27 ]. Notably, these studies did not employ gold‐standard methods for uterine fibroid detection and do not represent the true effect of fibroids on ASCVD. Although the methodological limitations of these studies prevent a causal interpretation, the upper (1.8) and lower (1.3) bounds were used as relative risk (RR) estimates for hypertension in this simulation for pedagogy. For simplicity, we focused on fibroids as the exposure, though similar concerns apply when fibroids are the outcome. Sample size and parameters: We fixed the sample size to 100,000 females and considered varying prevalences (5%, 20% and 60%) of uterine fibroids based on a community‐sampled cohort of reproductive‐aged females who underwent transvaginal ultrasonography [ 2 ]. Of note, we used a range of estimates reported in this study in their stratified analyses by age group, as they excluded those with current hormone use or with prior ovarian surgery, which may underestimate prevalence. Sensitivity, denoted as L , ranged from 0% to 100%. Specificity, denoted as M , was fixed at 90%, as an approximate average of the ascertainment methods. Variables for subsequent equations are defined in Table  2 . Contingency table defining outcome of hypertension by uterine fibroid status. Equations for RR were defined as: (1) True RR = A A + B C C + D (2) Each of the following examples varied additional specific parameters as outlined below. The observed RR based on the simulated data were calculated across 0%–100% sensitivity and compared to the fixed true RR (1.8, 1.3). [Correction added on 11 September 2025, after first online publication: Equation 2 has been corrected in this version.] In exploratory analyses, we considered how detection bias, or more specifically differential misclassification with respect to the outcome, affects estimates, given that those with hypertension may have higher fibroid detection due to increased surveillance, particularly in cross‐sectional studies. Sensitivity was increased by 20% for those with hypertension and true RR of 1.0, 1.3 and 1.8 were considered.

We assessed how varying fibroid sensitivity affected observed hypertension risk ratio estimates. Figure  1A,B shows observed versus true RRs (1.8 and 1.3) across sensitivities, with dashed lines marking sensitivities of standard methods. Simulated observed risk ratios compared to the true risk ratio of uterine fibroids on the risk of hypertension by varying the sensitivity of identifying uterine fibroids and population prevalence, N  = 100,000. Prevalence of uterine fibroids refers to the overall population prevalence. Specificity was fixed at 90%. Regardless of prevalence, decreasing sensitivity leads to a bias towards the null, and at very low sensitivity (10%), the observed RR can cross the null (Figure  1 ). The observed RR was furthest from the true RR at ascertainment methods with low sensitivity (self‐report) and closest to the true RR with high sensitivity (transvaginal ultrasonography). With perfect sensitivity, the observed RR was further from the true RR when the prevalence of uterine fibroids was low compared to high. This suggests that if the population prevalence of uterine fibroids was low (5%), implementing the gold standard for ascertainment of fibroid status will bias results, due to imperfect specificity. This is less of a concern if the population prevalence of uterine fibroids is assumed to be high (60%). Based on these findings, the reported RR in the original fibroids and hypertension studies was likely to be an underestimate of the true RR. Results are shifted towards the null at lower sensitivities when the true RR for uterine fibroids and hypertension was 1.3 (Figure  1 ). In exploratory analyses examining detection bias (differential misclassification of fibroids with respect to the outcome of hypertension), results for true effect estimates of 1.8 and 1.3 are similar to the primary analysis (Figure  S1 ). However, if the true effect estimate was null (1.0), then detection bias will lead to bias away from the null; that is, researchers will identify a positive association between exposure and outcome even in scenarios where one does not exist. Next, we consider how fibroid symptoms affect results. In truth, yet often unknown to investigators, there are three groups: females with (1) symptomatic fibroids, (2) asymptomatic fibroids and (3) no fibroids. Of note, symptom status is not only related to size, but also location [ 1 , 26 ]; small fibroids that are submucosal and bulge into the uterine cavity can cause heavy menstrual bleeding, whereas large, subserosal fibroids that grow outside of the uterus may be asymptomatic. We acknowledge that categorising females with uterine fibroids as symptomatic or asymptomatic may overlook those who experience symptoms for years before seeking clinical care and receiving a diagnosis, as the time to seeking care is often unmeasured. Two sensitivity parameters are now considered: that of symptomatic fibroids and that of asymptomatic fibroids: (3) L a = Sensitivity = Probability Identifying uterine fibroids Females with asymptomatic uterine fibroids ] (4) L s = Sensitivity = Probability Identifying uterine fibroids Females with symptomatic uterine fibroids ] In observed data analysed by investigators, there are two groups: females identified (1) with fibroids; (2) with no fibroids. The proportion of misclassified females will depend on the sensitivity and specificity of the ascertainment method (Table  S1 ). The impact of differential sensitivity in detecting uterine fibroids by symptom status on results depends on the research question. Using hypertension as the outcome, there are four scenarios: (1) Asymptomatic fibroids have no effect on hypertension, whereas symptomatic fibroids do have an effect. (2) Asymptomatic and symptomatic fibroids have the same effect on hypertension. (3) A dose–response relationship exists, where more severe symptoms increase the effect on hypertension risk. (4) Asymptomatic fibroids are different from symptomatic fibroids and subsequently may have a larger effect on hypertension; this could be due to distinct physiologic features, a tendency to be located in specific areas, or that they go untreated, leading to greater long‐term consequences [ 28 ]. Our second simulation considers these four scenarios outlined in Table  3 . Definitions of scenarios and associated parameters considered for simulation study. Abbreviation: RR, risk ratio. Sensitivity: L a = probability of classifying uterine fibroids as having uterine fibroids among those who have asymptomatic uterine fibroids. Again, we fix specificity at 90% and consider three population prevalences of uterine fibroids among a sample size of 100,000 females. In this case, we assume that 75% of uterine fibroids are asymptomatic, whereas 25% are symptomatic [ 26 ]. In two exploratory analyses, we repeat the simulation, assuming 50% and 25% are asymptomatic, respectively. For simplicity, we fixed sensitivity for symptomatic fibroids ( L s ) at 85%, as an average of the lower bounds reported in validation studies. The sensitivity for asymptomatic fibroids ( L a ) varied from 0% to 100%. For each case, we set the true RR based on the four assumptions: (5) True RR symptomatic vs . no fibroids = A A + B E E + F (6) True RR asymptomatic vs . no fibroids = C C + D E E + F and calculated the observed RR, comparing uterine fibroids to no uterine fibroids, which includes those with no fibroids as well as those with asymptomatic fibroids, dependent on the sensitivity and specificity (fixed at 90%) of the ascertainment method: (7) Observed RR fibroids vs . no fibroids = L s A + L a C + 1 − M E L s A + L a C + 1 − M E + L s B + L a D + 1 − M F 1 − L s A + 1 − L a C + ME 1 − L s A + 1 − L a C + ME + 1 − L s B + 1 − L a D + MF Figure  2 displays the observed RR, compared to the true RR, of symptomatic and asymptomatic uterine fibroids, respectively, compared to no uterine fibroids, with varying sensitivity of identifying asymptomatic uterine fibroids for each scenario outlined in Table  3 . [Correction added on 11 September 2025, after first online publication: Equation 7 has been corrected in this version.] Simulated observed risk ratios of uterine fibroids, overall, on risk of hypertension by varying sensitivity of asymptomatic identifying uterine fibroids, N  = 100,000. Prevalence of uterine fibroids refers to the overall population prevalence. Specificity was fixed at 90%. Sensitivity of detecting symptomatic uterine fibroids was fixed at 85%. In Scenario 1, assuming that only symptomatic uterine fibroids affect hypertension, decreasing sensitivity of detecting asymptomatic uterine fibroids shifts the observed RR away from the null by capturing more asymptomatic cases (Figure  2 ). At both high (100%) and low (0%) sensitivity, the observed RR does not reach the true RR of symptomatic vs. no fibroids nor asymptomatic vs. no fibroids. When assuming a population prevalence of uterine fibroids at 5% and Scenario 1, the observed RR approximates the true RR of asymptomatic vs. no fibroids at all sensitivities and never approximates the true RR of symptomatic vs. no fibroids. Note that Scenario 2 utilises the same parameters as Example 1, though the sensitivity of symptomatic fibroids was fixed at 85%. Increasing sensitivity shifts the observed RR towards the true estimate. At very low sensitivity, the observed RR was the most biased and approximates the null, but unlike Example 1, it never crosses the null. When assuming a dose–response relationship in Scenario 3, the observed RR was attenuated towards the null but approached the true RR of asymptomatic versus no fibroids with increasing sensitivity. Scenario 4, with the assumption that asymptomatic fibroids have a stronger effect on hypertension than symptomatic, demonstrates that when sensitivity was low, the observed RR was attenuated beyond the true RR of symptomatic vs. no fibroids but shifts towards the true RR of asymptomatic vs. no fibroids with increasing sensitivity. At lower prevalence of uterine fibroids, the observed RR in Scenario 3 and 4 was attenuated further from the true RR of asymptomatic vs. no fibroids. In the case of fibroids and hypertension, Scenario 1 may have highest biologically plausibility, since chronic inflammation may be greatest with the most severe fibroids [ 11 ]. Therefore, as Example 1, the reported RR in the original fibroids and hypertension studies reliance on self‐report was likely to be an underestimate of the true RR if the true population prevalence of fibroids was less than 20% but close to the truth at 60% prevalence. Table  2 displays alternative scenarios assuming a true modest effect size (1.3). Results do not change appreciably when assuming a modest true RR (Figure  S2 ). Exploratory analyses (Figures  S3–S6 ) showed that the results were most biased when 75% of fibroid cases were asymptomatic and least biased when asymptomatic cases accounted for only 25% of fibroid cases. Myers et al. compared self‐reported vs. ultrasound‐detected fibroid status (ultrasound as the gold standard) among Black and White females in the RFTS and UFS studies [ 19 ]. Sensitivity of self‐reports ranged from 15% to 79%, varying by fibroid size, race and age, with 50%–80% higher sensitivity among Black females. Due to limited data on other racialised groups, Example 3 focuses on differential misclassification among Black and White females, though similar issues likely affect other groups. We assumed that among those who have uterine fibroids, 75% would be of Black racialised identity and 25% White, to reflect that Black females have 3 times the risk of developing uterine fibroids compared to White females [ 2 ]. For example, when the population prevalence of uterine fibroids was set to 20% among a sample of 100,000, of the 20,000 that are set to have uterine fibroids, 15,000 (75%) are assumed to be Black females and 5000 (25%) White females. The racial distribution of the cohort without uterine fibroids was assumed to be 15% Black and 85% White to reflect US Census estimates [ 29 ]. Sensitivity was set to 50% higher among Black females than White females; for instance, when sensitivity was 50% for White females, it was set to 75% for Black females. Figure  3 displays the observed RR, compared to the true RR of uterine fibroids compared to no uterine fibroids, with varying sensitivity overall and by race. In all cases, deviation from the true RR was greatest with low sensitivity ascertainment methods, specifically the upper and lower bounds of self‐report. The overall and stratified results among Black females are similar, but the stratified results among White females are attenuated towards the null at all sensitivities. At higher prevalence when sensitivity was < 50%, the results for Black females are more biased compared to White females. Simulated observed risk ratios of uterine fibroids on risk of hypertension with varying sensitivity overall and by race, assuming uterine fibroid prevalence of 20% and a true risk ratio of 1.8, N  = 100,000. Prevalence of uterine fibroids refers to the overall population prevalence. Specificity was fixed at 90%. Results do not change appreciably when considering a modest true RR (Figure  S7 ). To illustrate how a sensitivity analysis can address measurement error, we repeated the primary example, assuming a population prevalence of 20% and applying a quantitative bias analysis (QBA) to adjust results. QBA methods aim to quantify the influence of systematic error on an epidemiology study's estimate of effect [ 30 ]. We assumed that fibroid data were ascertained via ICD codes in this example. Thus, we applied the sensitivity and specificity parameters from Table  1 using the ‘episensr’ package in RStudio: https://dhaine.github.io/episensr/articles/episensr.html . Figure  S8 shows the results. At all sensitivities, the corrected RR was closer to the true RR than the observed RR. At the parameters specified in the QBA (sensitivity: 82.7%, specificity: 96.6%), the corrected RR (1.70) closely approaches the true RR (1.80); however, due to differing specificity, it does not reach the true estimate.

Uterine fibroids are a common gynecologic condition associated with significant morbidity, including heavy menstrual bleeding, pelvic and back pain, and anaemia [ 1 ]. Between 25% and 35% of individuals with a uterus assigned female at birth (hereafter females) are diagnosed with uterine fibroids during their reproductive‐aged years [ 1 , 2 ]. However, estimates suggest nearly 70% of females develop fibroids before menopause [ 1 , 2 , 3 , 4 ]. These non‐cancerous tumours and resulting treatments decrease quality of life [ 1 ]. Uterine fibroids have been linked to adverse health outcomes in some studies, including hypertension, a major contributor to cardiovascular deaths in the US [ 5 , 6 ]. Although correlations between uterine fibroids and hypertension have been observed, causality is unclear [ 7 , 8 , 9 , 10 ]. One proposed mechanism is that fibroids increase creatine kinase production, which may trigger vascular smooth muscle cell proliferation, elevating blood pressure [ 7 , 11 ]. Fibroids may also induce inflammation, causing metabolic changes and narrowed blood vessels [ 7 ]. Given hypertension's role in cardiovascular disease, understanding whether uterine fibroids can be targeted to reduce disease burden has significant public health implications [ 12 ]. To accurately assess the association between fibroids and hypertension, rigorous, unbiased studies are needed. Clarifying this relationship could inform interventions, such as enhanced fibroid and hypertension screening; however, the lack of available and effective treatment options for fibroids may limit future interventions. Of note, this importance is relevant for all outcomes considered to be linked to the long‐term health implications of uterine fibroids, such as depression and anxiety [ 5 ]. Diagnosing uterine fibroids is challenging, leading to inconsistent epidemiologic data [ 6 ]. Symptoms of heavy menstrual bleeding and pain may be dismissed by both patients and providers as ‘normal’ menstrual experiences [ 13 , 14 ]. Further, nearly 30% of reproductive‐aged females do not engage in regular reproductive healthcare due to access barriers, reducing diagnostic opportunities [ 15 ]. Additionally, gold‐standard diagnostic methods (i.e., pelvic MRIs, transvaginal ultrasounds) are expensive and often inaccessible [ 16 ], leaving diagnoses reliant on pelvic exams prone to error. These factors contribute to underdiagnosis and misclassification, a concern in epidemiologic research. Misclassification is further influenced by two key factors. Approximately 75% of fibroid cases are asymptomatic [ 1 ], leaving many undiagnosed. Second, diagnosis and misclassification may vary by racialised group. Black females have 2–3 times higher prevalence than White females using transvaginal ultrasonography and higher self‐reported sensitivity [ 1 , 17 , 18 , 19 ]. Structural racism also affects access and quality of care, potentially leading to differential misclassification when relying on self‐reports or clinical diagnoses [ 17 ]. Recognising these disparities and the role of symptom status is crucial when evaluating how underdiagnosis biases epidemiologic findings. This simulation examines how varying sensitivity in detecting uterine fibroids and resulting misclassification affect effect estimates, using the association between fibroids and hypertension onset as a motivating example. We address three exposure misclassification scenarios: (1) varying fibroid detection sensitivity, (2) varying fibroid detection sensitivity by symptom status and (3) varying fibroid detection sensitivity by race, reflecting disparities in diagnosis.

Researchers should use highly sensitive methods and include quantitative bias analyses to reduce bias in fibroid studies. Future validation studies are vital to improving accuracy and ensuring findings are robust and generalisable, benefiting research on gynecologic conditions and health disparities.

The authors declare no conflicts of interest.

Data S1.