Recall bias in population-based case-control studies of ovarian cancer and genital talcum powder use: potential impact and quantitative bias analysis

Abstract

Introduction

Quantitative bias analysis (QBA) methods have been developed, with growing applications to epidemiological studies. Bayesian QBA (BQBA) methods may be superior in effectively using evidence-based priors and study data to displace fixed-value assumptions, and in characterizing distributions of adjusted point estimates and associated uncertainty.

Methods

We applied simple QBA (SQBA) and BQBA methods to population-based case-control studies on ovarian cancer and genital talcum powder use. Our hypothesis was that differential misclassification of self-reported historical talcum powder use led to statistically significant associations in several of these studies. SQBA results covering ranges of published exposure sensitivity and specificity estimates for genital talcum powder use were generated. BQBA models generated posterior means and 95% credible intervals (CrI’s) of bias-adjusted odds ratios (ORs), sensitivities and specificities of controls and cases, respectively, using pooled case-control study data.

Results

The mean Bayesian posterior OR was 1.08 (95% CrI 0.67–1.77), mean case and control sensitivities were 0.87 and 0.78 and mean case and control specificities were 0.86 and 0.90, respectively. Model diagnostics indicated good fit.

Conclusion

Upon adjusting only for hypothesized differential exposure misclassification bias, the association between ovarian cancer and genital talcum powder use was attenuated relative to the unadjusted one. While other potential sources of bias may be operating within this body of studies, exposure reporting bias likely accounts for the statistically significant ORs reported in the individual case-control studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-026-02836-x.

Introduction

It is well understood that as sample size increases in epidemiological studies, the amount of random error is reduced. However, it appears less well understood that the amount of systematic error usually remains unchanged and may aggravate problems with interpretation [1]. Associations from epidemiological studies that have large sample sizes and relatively tight confidence intervals (CIs) around effect measure estimates present opportune situations to examine through quantitative bias analysis (QBA) the component of nonrandom (or systematic) error in the apparent association [2, 3] Various QBA approaches, methods and statistical tools to have evolved and are increasingly utilized to quantify and correct for the impact of various systematic errors, ultimately to assist in drawing inferences from epidemiological studies [2–5].

Examining and correcting the potential impact of exposure recall and reporting bias and subsequent exposure misclassification on reported odds ratios (ORs) is an example where the value of QBA has been demonstrated [6–9]. In case-control studies, recall bias may occur when cases have a greater ability than controls to recall and accurately report historical exposures or risk behaviors. More technically, recall bias can arise when the threshold for reporting exposure is different in cases and controls, frequently seen and characterized as greater sensitivity of the cases’ recall and estimation of historical exposure(s) compared with that of controls, as well as greater specificity of controls’ recall and estimation of non-exposure(s) compared with that of cases. This is one of the main reasons why epidemiologists seek to match cases and controls on general health status and recent encounters with healthcare systems (e.g., as in hospital-based case-control studies) and to avoid relying on self-reported exposures in favor of more objective data sources.

Recall bias has been documented in exposure validation studies that used medical records to validate self-reported exposure or other risk factors, although the patterns of misclassification can vary. For example, Drews, Kraus and Greenland [10] in a case-control study of mothers of sudden infant death syndrome (SIDS) patients (cases) compared to living controls found a typical recall bias pattern for 18 of the 25 variables examined. Schwartz et al. [11] found differences between reporting of previous flu vaccination in people with cancer (i.e., cases) compared to people without cancer or other serious flu infection risk factors (i.e., controls) (71.5% vs. 60.5% for sensitivity and 81.6% vs. 90.0% for specificity, respectively) and sensitivity was much poorer in those under age 65 (46.9% in cases vs. 29.1% in controls). Vestergaard et al. [12] found that breast cancer patients had 86.2% sensitivity compared to 80.6% in controls when self-reporting shift work, while controls had a specificity of 83.7% compared to 82.6% in patients. The authors reported that the OR in a “hypothetical population” would be reduced to 1.05 (95% simulation interval (SI): 0.95–1.16) from an observed OR of 1.12 (95 CI: 1.03–1.21) when adjusting for recall bias. However, in a letter to the editor, Burstyn and Luta [13] commented that Vestergaard et al. [12] (a) made an error in their SQBA such that the reported 95% credible intervals (CrI’s) of 0.88 to 1.27 should have been wider; and (b) that the Bayesian method, which additionally accounted for uncertainty in misclassification parameters, yielded an adjusted OR with mean of 0.98 and 95%CrI of 0.3–1.7. This highlights that applying the more rigorous method generated interpretationally different results and thus is preferred.

Ovarian cancer is a serious cancer with a high case-fatality rate and limited preventive strategies [14] Several risk factors have been identified – some positively and some negatively associated with ovarian cancer risk, including genital application of talcum powder [15, 16]. An association between ovarian cancer and self-reported historical genital talcum powder use was first reported in 1982 [17] and subsequently in numerous additional population-based case-control studies [18–22] Kadry Taher et al. [20] conducted meta-analyses of three cohort and 24 case-control studies spanning four decades (1982 to 2016) and including 16,005 ovarian cancer cases and 201,881 controls. The meta-OR for population-based case-control studies was 1.34 (95% CI: 1.27–1.41) but unremarkable for the cohort (OR = 1.06, 95% CI: 0.90–1.25) and the hospital-based case-control (OR = 0.96, 95% CI: 0.78–1.17) studies. Nevertheless, the observed association in the population-based case-control studies rapidly became the impetus for major toxic tort litigation, primarily in the US, and driver of product safety regulatory actions worldwide.

For these reasons we believe that this topic provides fertile ground for conducting a QBA for possible differential exposure reporting bias and subsequent misclassification. In this paper, we examine the potential role of differential exposure misclassification based on published results from population-based case-control studies of the relationship between ovarian cancer and use of genital talcum powder. Our specific objectives for this paper were (1) to visually describe the range of observed ORs when varying self-reported exposure estimate sensitivity and specificity levels in cases and controls; (2) to present a simple quantitative bias analysis (SQBA) where modest differential decreases in sensitivity and specificity are assumed in the reporting of genital talcum powder use between ovarian cancer cases and controls; and (3) to perform a Bayesian quantitative bias analysis (BQBA) on these population-based case-control studies to evaluate the change in OR and uncertainty due to exposure recall bias. In this study we do not, however, assume the presence or absence of, or address, potential confounding bias, which would require additional sophisticated analyses coupled with additional assumptions regarding the direction and magnitude of uncontrolled confounding bias. We end with discussing differences in the methods, advantages and disadvantages, and implications for interpreting observed results from the literature on ovarian cancer and use of genital talcum powder.

Methods

Available QBA tools range from simple sensitivity testing of single parameters to sophisticated Bayesian approaches. SQBA essentially draws on validation information or educated guesses if validation against a gold standard is unavailable regarding the potential direction and quantity of systematic error in a key parameter such as exposure, outcome or potential confounding factors. These fixed values are then applied to derive “bias-adjusted” relative risk estimates. Analyses may be repeated using other bias parameters to obtain additional bias-adjusted estimates of association, but without reliable variability estimates or confidence intervals as the assumed input values are fixed. The use of sensitivity, specificity, predictive value or other indicators of direction and degree of systematic error is common, whether derived from a validation exercise or assumed. The main concern about SQBA, despite their straightforwardness, is that small deviations in fixed values of assumed bias parameters can have unexpectable large impact on the calculations and thus are not robust to misspecification of the bias parameters [5]. In contrast, Bayesian adjustments are considered superior and standardized methods and computer codes for these are more widely available [2–4]. BQBA approaches have become favored because they not only derive estimates of distributions of true values of one or more parameters (given data, models and priors), but also validly address uncertainty. Below we illustrate both.

Selection of case-control study data

To simplify the QBA, we combined the published study data from other pooled population-based case-control studies instead of performing QBA on several smaller individual studies. Specifically, Boon et al., [18] the most recent systematic review, identified six pooled case-control studies evaluating ovarian cancer and genital talcum powder use [23–28]. The six pooled case-control studies were reviewed to determine the degree of possible overlap among the original studies. Pooling all reported data from Davis et al., [23] Phung et al., [24] and Cramer and Xu [28] with the Southern Ontario Ovarian Cancer Study data from Terry et al. [26] and the Nurses Health Study (NHS) nested case-control data from Gates et al. [27] constitutes an apparently non-overlapping pooled dataset of 7,705 cases and 15,009 controls. The pooled crude OR for these studies is 1.11 (95% CI: 1.05–1.18). To more closely match the data used in the meta-analyses, we removed the nested case-control study data (NHS in Gates et al. [27] and Women’s Health Initiative (WHI) in Davis et al. [23]) as these should not be subject to the same hypothesized biases arising from post-diagnosis recall and reporting of historical genital talcum powder use. The final data set of pooled population-based case-control studies contained 6,997 cases and 10,965 controls (Table 1). The resulting crude OR for genital talcum powder use and ovarian cancer in this data set is 1.32 (95% CI: 1.24–1.41).

Table 1.

Case-control studies selected for quantitative bias analysis (see text for details)

	cases talc +	cases talc -	controls-talc +	controls talc -	Reported OR (95% CI)	Crude Calculated OR (95% CI)*	Reported Genital Talc Use
Davis 2021 (without WHI nested c-c)	753	2047	864	3213	1.34 (1.01–1.79) 1.31 (1.15–1.48)	1.37 (1.22–1.53)	Combined using ≤ 1/wk and > 1/wk
Phung 2022 (endo +)	79	220	106	323	1.38 (1.04–1.84)	1.13 (0.80–1.58)	use of talc
Phung 2022 (endo -)	827	2172	1304	4137	1.12 (1.01–1.25)	1.21 (1.09–1.34)	use of talc
SON from Terry 2013*	197	252	200	364	1.35 (1.03–1.76)	1.42 (1.10–1.83)	ever regularly used talc powder
Cramer 1995	201	249	154	300	1.6 (1.2–2.1)	1.57 (1.20–2.06)	used talc
POOLED TOTALS	2,057	4,940	2,628	8,337	- - - -	1.32 (1.24–1.41)

* SON Southern Ontario Ovarian Cancer Study

Simulation study: variability in ORs with different combinations of sensitivity and specificity of exposure classification

Simulation study of the impact of misclassification of exposure requires fixing realistic values of the prevalence of the risk factor, i.e., genital talcum powder use. In a pooled analysis of eight population-based case-control studies, Terry et al. [26] reported 31% prevalence among cases (range across studies was 15% to 49%) versus 25% among controls (range across studies was 15% to 45%). These reported prevalence ranges are used to bound reasonable prevalence ranges in the simulations below.

To investigate the impact of exposure misclassification on OR estimates, we conducted a simulation study exploring how varying exposure estimates of sensitivity and specificity levels affect the observed OR () relative to the true OR (). The simulation quantifies how imperfect (i.e., less than 100%) sensitivity and specificity of the individual exposure estimates affect the relationship between observed and true ORs.

Let and be the exposure estimate sensitivity for cases and controls, and be specificity for cases and controls. The true cell counts can be derived from the observed counts through the following equations:

Note that: A = exposed cases, B = non-exposed cases, C = exposed controls, D = non-exposed controls, a = observed exposed cases, b = observed non-exposed cases, c = observed exposed controls, d = observed non-exposed controls.

We used a total cohort size of 4000 with half designated as cases and half as controls. We then calculated the true counts A, B, C, D based on the presumed true exposure prevalence among cases () and set the true OR at 1, 3, and 5, respectively. The observed ORs were computed with different levels of and . We set to range from 0.6 to 1.0, and to be from 0.6 to 0.95, increasing in increments of 0.01, and was set at 0.85, 0.9, 0.95 and 0.99, and at 0.9, 0.95 and 0.99. Table 2 summarizes the true contingency table under each example that we have examined.

Table 2.

True contingency tables for examples with different and exposure prevalence among cases

Example		Exposure prevalence among cases ()		Exposed	Non-exposed
1	1	0.5	Cases	1000	1000
			Controls	1000	1000
2	1	0.3	Cases	600	1400
			Controls	600	1400
3	3	0.3	Cases	600	1400
			Controls	250	1750
4	5	0.3	Cases	600	1400
			Controls	158	1842

For each parameter combination, we computed the observed counts (a, b, c, d) given the true counts (A, B, C, D) and the specified values for sensitivity and specificity, which in turn allowed us to calculate the observed .The difference between and was recorded. Heatmaps were generated to visualize the magnitude and direction of bias across sensitivity-specificity combinations, highlighting patterns such as systematic inflation or attenuation of the relative to the true value.

SQBA approach

To evaluate how changes in sensitivity and specificity of the exposure estimates could affect the derived from observed study data (i.e., 1.32), we generated heatmaps for our pooled case-control study data examining the association between ovarian cancer and self-reported historical genital talcum powder use. We applied the method described above using the observed contingency table rather than simulated data and computed the bias-adjusted accounting for hypothesized exposure misclassification due to recall bias. Distinct ranges for and were specified to reflect expert recommendations summarized in Table A2.1 of the International Agency for Research on Cancer (IARC) Monograph 136 [16]. Specifically, we set both and to range from 0.6 to 1 in increments of 0.01, while setting to 0.85, 0.9, and 0.95, and to 0.8, 0.85, 0.9 and 0.95, respectively. These ranges align well with the experts’ opinions, which suggest values between 0.6 and 0.95, between 0.65 and 0.98, between 0.85 and 0.95, and between 0.8 and 0.95. Differential values for sensitivity and specificity among cases and controls were allowed, as this was central to our hypothesis regarding reporting bias leading to exposure misclassification. As with other standard applications of SQBA, uncertainty around sensitivity and specificity cannot be addressed and therefore was not formally incorporated here. We were particularly interested in the consequences of SQBA using ranges of misclassification parameters elicited from three experts who contributed to the SQBA presented in IARC Monograph 136 [16].

BQBA approach

We conducted our formal BQBA using methods to account for uncertainty in exposure misclassification – one distinct advantage of BQBA – in the case-control studies of ovarian cancer and self-reported historical genital talcum powder use. Unlike SQBA, which relies on fixed values of sensitivity and specificity, the Bayesian framework incorporates prior distributions for these parameters and formally propagates their uncertainty into posterior estimates including the ORs. Burstyn et al. [29] illustrated a Bayesian framework to estimate bias-adjusted, true based on inputs of several parameters including exposure prevalence, sensitivity, and specificity, following the work of Gustafson [30]. The Bayesian method starts with specifying priors for each parameter. Posterior inference was carried out using Markov Chain Monte Carlo (MCMC) sampling with Gibbs sampler, yielding posterior distributions for , and that of all of the model’s parameters. The resulting posterior distribution is a combination of priors and the given data. We follow the textbook implementation of Gustafson [30].

Specifically, let denote total number of cases, and total number of controls. Given the probability of cases reporting being exposed, and the same for controls as , and can be modeled from binomial distribution with: . Letting and be the true exposure prevalence in controls and cases, allowing differential misclassification, we have:

The case prevalence is linked to the true OR, , by:

We place Beta-distributed priors on the control prevalence , and as they are all bounded between 0 and 1. The scale and shape parameter for sensitivities and specificities were based on means of “best guess” and 2.5% lower quantiles of “95% range” elicited from the three experts’ opinion in Table A2.1 of IARC Monograph 136 [16] (see Appendix 1 for details of derivation of these prior parameters in R; Fig. 1S in Appendix 1 illustrates the shapes of the density of these priors ). We set the prior on the log of , which is a vague and null-centered prior that approximates a normal distribution with mean 0 and variance 0.75, i.e. 95%CI of OR from 0.1 to 10 [31]. We used the uninformative prior for the prevalence of exposure among control, which equals likelihood of values between 0 and 1. R code that we used to implement Bayesian adjustment is provided as Appendix 2. Convergence is assessed using trace plots, effective sample size, and (target 1.01).

Heuristically, in the Bayesian adjustment, we sample two candidate values of (, , ) from their respective priors and use them to calculate two candidate sets of values (, . The likelihoods of c and a are then evaluated using binominal distributions of (, and these likelihoods are combined with the likelihoods of the candidate values (, , , ) determined by their distributions. The candidate values of parameters of interest, such as , are preferentially retained if they have higher total likelihoods among the two candidate sets and are deemed to be samples from the posterior distribution. The retained sample from the posterior is then compared to the next set of candidate values and these are preferentially retained if they either improve or do not degrade the likelihood. The iterative procedure is repeated many times to obtain a stable sample from the posterior. This is different from multiple SQBAs in which (, ) are drawn systematically from a range of values and are combined with the observed data (a, b, c, d) to calculate , without evaluation of the likelihood, such that e.g., adjustments with extremely unlikely values of the misclassification parameters are given the same importance as those that are based on the “best guess” of the misclassification parameters.

Prior sensitivity analysis

In our default approach (scenario 1), we elicited prior distributions of misclassification parameters from the IARC experts’ best guesses using an averaging approach that minimized the influence of any single expert, while treating the collective informativeness of the experts’ judgements as reasonable. Prior sample sizes differed by parameters of prior distributions (~ 200–300, equivalent to claiming that the experts collectively had the knowledge that they derived from a validation study of 200 to 300 individuals). That is, we considered that best guesses by the experts are the most trustworthy, by definition. We conducted multiple sensitivity analyses to evaluate these assumptions: that informativeness of average of such best guesses can be weaker (scenarios 2 and 3) or stronger (scenario 4) than that believed by IARC’s experts. Specifically, in scenarios 2 to 4, we fixed mode of beta-distributed priors on misclassification parameters at the averages of best guesses of the three experts. However, we associated these modes with either lower precision than in scenario 1 for scenarios 2 and 3, with prior sample sizes fixed at 10 and 40, respectively, or greater precision than in scenario 1 with prior sample size of 400 for scenario 4. We further examined the impact on our results of believing that one of the experts should be trusted while the others are not trustworthy. We accomplished this in scenarios 5 to 7, by eliciting priors from best and lower value guesses of only one expert, from expert 1 to 3, respectively. Lastly, we revisited a key decision in how we constructed priors in our default approach, scenario 1. Instead of using the means of the lowest values provided by the IARC’s experts to indicate that 95% of prior densities take on greater values, we used the means of the highest values provided by the IARC’s experts to indicate that 95% of prior densities should take on lesser values.

Results

Potential impact of differential exposure misclassification

The heatmaps in Fig. 1 illustrate the spectrum of how differential misclassification can bias the relative to the for a range of plausible estimates of sensitivity, specificity and prevalence of historical genital talcum powder use. Each bottom x-axis and left y-axis represents control sensitivity () and case sensitivity (), respectively. The top x-axis represents case specificity (), while the right y-axis shows control specificity (). Each cell is color-coded to show the value of , with red indicating inflation bias and blue indicating attenuation bias. The heatmap reveals that bias leading to inflation of the is most prominent when both case and control exposure estimate specificity values are low. When specificities are held constant, inflated ORs tend to occur under low control sensitivity and high case sensitivity. The degree of inflation is also influenced by the : smaller result in more widespread bias leading to OR inflation. Additionally, exposure prevalence plays a role: higher prevalence leads to more inflation in the upper right region (high sensitivity and specificity), while the lower left region shows less inflation.

Fig. 1.

Heatmap visualization of difference between observed OR and true OR for example 1 (a), example 2 (b), example 3 (c), and example 4 (d)

Results of Simple Quantitative Bias Analysis (SQBA)

Figure 2 presents the results of the SQBA for the pooled population-based case-control data on talcum powder use and ovarian cancer. Each heatmap displays the difference between bias-adjusted OR, , and observed OR, , across the specified ranges of sensitivity and specificity values. Consistent with the simulation study, case sensitivity and control sensitivity are displayed on the bottom x-axis and left y-axis, respectively, while case specificity and control specificity are shown on the top x-axis and right y-axis.

Fig. 2.

Heatmap results for the pooled data from a SQBA

As expected, the direction of adjustment varied substantially with changes in sensitivity and specificity. Several trends are apparent from the heatmaps. First, when is fixed, the tends to be deflated as increases, demonstrated by expansion of blue regions from top to bottom. On the other hand, when is given, the tends to inflate with higher values of , indicated by the expansion of red area horizontally. In addition, when both and are fixed, deflated values are most evident in the upper left corners of the heatmaps, corresponding to situations where case sensitivity exceeds control sensitivity .

In this example, when is at approximately 0.85, inflated are likely to be observed, except when is at the relatively low level of 0.8. While when and is between 0.8 and 0.9, the is likely to fall below , especially when , i.e. the observed OR is null or greater than 1.

When we focus on the impact of SQBA in the ranges of misclassification parameters consistent with those elicited from three experts who contributed to IARC Monograph 136 [15], we notice that qualitatively different adjustments are obtained depending on the choice of experts (Fig. 3). The left-hand panel of Fig. 3 is our best approximation of the misclassification parameters in the IARC Monograph [15], with specificity of cases fixed at 0.85 and specificity of controls fixed at 0.9. Depending on the combinations of sensitivity for cases and controls, ORs are adjusted above or below the observed values. The proportion of OR adjusted up or down depends on the choice of the expert. This lack of robustness is one of the primary motivations for Bayesian adjustment for recall bias, results of which are presented below.

Fig. 3.

Heatmap results using experts’ suggestions on sensitivities and specificities; specificity among controls is 0.9

Results of Bayesian Quantitative Bias Analysis (BQBA)

Table 3 summarizes the posterior means, 95% CrIs, and selected model diagnostics from the Bayesian quantitative bias analysis (BQBA) using the pooled case-control data. The posterior mean of the bias-adjusted OR is 1.08 (95% CrI: 0.67–1.77), with the interval firmly embracing the null value. This indicates that the true association between ovarian cancer and genital talcum powder use likely is much weaker than the meta-ORs reported in the meta-analyses and those reported in the individual studies that were pooled for this analysis. The reduction in the mean of adjusted OR can be attributed to differential exposure misclassification between cases and controls.

Table 3.

Posterior means, 95% credible interval (CrI) and selected model diagnostics of bias-adjusted, true, odds ratio , sensitivities (p) and specificities (q), as well as prevalences ( r ) of controls (subscript 0) and cases (subscript 1) using pooled data. = 1 for all parameters

Parameter: case status			Posterior mean	95% CrI	Effective sample size
True odds ratio			1.08	0.67–1.77	18,815
Sensitivity	controls		0.78	0.73–0.83	262,745
	cases		0.87	0.81–0.91	455,187
Specificity	controls		0.90	0.85–0.93	19,729
	cases		0.86	0.81–0.90	55,653
True prevalence of exposure	controls		0.20	0.14–0.25	19,255
	cases		0.21	0.16–0.26	57,077

Posterior distributions for the misclassification parameters were like the priors, indicating little learning about misclassification parameters. The mean was 0.78 (95% CrI: 0.73–0.83), while was slightly higher at 0.87 (95% CrI: 0.81–0.91). The and had posterior means of 0.90 (95% CrI: 0.85–0.93 and 0.86 (95% CrI: 0.81–0.90), respectively. These results indicate some differential misclassification across cases and controls.

Convergence diagnostics indicated that all parameters were estimated reliably. The were equal to 1 for all monitored parameters, and effective sample sizes were considerably high. The trace plots for, , r₀ and r₁ exhibit good mixing and stationarity across four chains, further supporting convergence. Details are provided in Appendix 3.

Additional analyses were performed as sensitivity assessments. Figure 4 summarizes the posterior distributions of the ORs obtained under different assumptions about the degree of exposure misclassification. Scenarios 2 and 3 assume weaker priors than in scenario 1 (detailed in Table 3) and reflect more uncertainty in the inference while the stronger prior of scenario 4 leads to a more concentrated posterior of ORs than scenario 1. The choice of an expert whose knowledge of exposure misclassification is used in the adjustment, i.e. scenarios 5 and 6 but does not alter overall conclusion about evidence for the existence of the association. However, different experts’ beliefs shifted adjusted effect estimates either above (scenario 5 and 6) or below the null (scenario 7), on average. Comparison of different ways to summarize the beliefs of the three experts, i.e. scenarios 1 versus 8, suggest that such choice does not alter overall conclusions and we conclude that our default approach led to more informative estimates. Overall, prior sensitivity analysis reassured us that our conclusions were robust to a reasonable range of assumptions about the degree of misclassification of recall of genital talcum powder use.

Fig. 4.

Prior sensitivity analysis for misclassification-adjusted association (odds ratios) of ovarian cancer with report of ever having used genital talc in population-based case-control studies; see text for description of each scenario; scenario 1 corresponds to our main analysis presented in Table 3

To summarize, these results indicate that the BQBA provides stable posterior estimates of both the adjusted ORs and the underlying misclassification parameters, while propagating uncertainty appropriately.

Discussion

The BQBA generated a posterior mean OR of 1.08 (95% CrI 0.67–1.77) based on an effective sample size of 18,815. This indicates that upon adjusting only for differential exposure misclassification bias, no association between ovarian cancer and genital talcum powder use remained if one were to apply the criteria used by some to judge “significance” of statistical results based on whether 95% confidence interval includes the null. A more nuanced and sensible interpretation of our main result is that the true OR is consistent with both protective effects as well as excess risk associated with the exposure. It is therefore difficult to see how the population-based case control studies of ovarian cancer and genital talcum powder use can be viewed as informative about even the direction, let alone magnitude, of any effect. The Bayesian adjustment also produced posterior values for the sensitivity and specificity for both cases and controls, which can be used as priors in future research that utilizes similar exposure assessment methods as those deployed in the studies considered here.

Our QBA included calculations of possible bias arising from all combinations (within evidence-based and reasonable assumed ranges of values of exposure prevalence and sensitivity and specificity) of reported genital talcum powder use. These were presented as heatmaps demonstrating that under different conditions nondifferential exposure misclassification assessed using SQBA methods can result in over- as well as under-estimation of an underlying association. This underscores the importance of using methods that are robust to misspecification of the bias parameters. Our SQBA results for any specific combination of sensitivity and specificity confirmed lack of robustness. As such, we propose that our BQBA results more accurately reflect the true range of possible associations as compared to our SQBA and that of Goodman et al. [32] and IARC [16] which we believe do not represent the best practices in QBA.

According to PubMed statistics, nearly 1400 publications since 1965 have been tagged “bias analysis,” but only 161 associated with the keyword “quantitative bias analysis,” 126 of which (78.3%) also were labeled “epidemiology” (counts based on PubMed search conducted on September 10, 2025). Nearly two-thirds of these have been published only since 2000. Clearly, interest has surged, and QBA increasingly is an accepted standard tool, with guidance and standardization available for the simple and probabilistic QBA since about 2014 [33, 34] Solidifying these approaches in the epidemiological and biostatistical repertoire are several major textbooks [2–4, 30], and increasing numbers of original research articles and reviews in which QBA was applied. While broadly appropriate and applicable to identify and correct for the effects of any source(s) of bias in epidemiological studies, the need for QBA appears to be greater in some epidemiological research contexts in which potential biases specifically should be assessed quantitatively, i.e., beyond stating that various sources of such biases including confounding “cannot be ruled out.” Thus, QBA can provide affirmative evidence of important bias and furthermore quantify the amount, and, using more suitable methods, characterize and inherent uncertainties in the bias parameters.

For example, differential recall in self-reported historical exposure information of cases and controls is one often-cited source of bias and may be a major limitation in interpreting the results of certain case-control studies. The IARC publication on QBA [4] highlighted as a “key point” that “[r]ecall bias is not an inherent feature of case–control studies; for example, exposure estimation may be based on historical records (e.g. work history records) or biospecimens banked in the past.” On the other hand, the collection of published population-based case-control studies on ovarian cancer and genital talcum powder use – the focus of our evaluation – relied entirely on post-diagnosis self-reported use. Indeed, it was recognized more than a decade ago that bias analyses might not be needed “until research reports contemplate alternative hypotheses and draw inferences … and certainly once policy decisions are contemplated, bias quantification by simple bias modelling becomes essential and more complex modelling may also be needed” (p. 1975) [33]. Today, any systematic review observing statistical associations between some risk factor and disease of interest might wish to consider incorporating QBA into causal inferential deliberations. As a case in point, the European Chemicals Agency (ECHA) Risk Assessment Committee (RAC) [35] recommended classifying talc as a CARC 1b carcinogen – presumed to be carcinogenic to humans – largely on the reported results of the population-based case-control studies of ovarian cancer and genital talcum powder use we used our study. This underscores the need to more fully understand the extent to which recall and reporting biases likely influenced the reported study results – and subsequently the regulatory decision. Interestingly, IARC Monograph 136 [15] included a SQBA that indicated that these studies indeed were likely susceptible to reporting bias and subsequent differential exposure misclassification as well as confounding bias. IARC however also classified talc as a “possible” human carcinogen, based in part on their SQBA for confounding and non-differential exposure misclassification in the cohort studies – which the RAC largely considered null [15].

Our BQBA further was motivated by the observation of associations in some of the case-control studies suggesting that the association might not entirely reflect an underlying causal association. First, among many of the studies analyzing exposure parameters beyond “ever” versus “never” genital talcum powder use indicated no clear relationship between ovarian cancer and the duration or frequency of genital talcum powder use, possibly indicating that users might differ from non-users on other risk factors for ovarian cancer. Second, some specific results appeared puzzling. Schildkraut et al. [36] examined several dimensions of self-reported talcum powder use and presented numerous ORs for different subgroups. ORs for self-reported “only nongenital use” were comparably associated with ovarian cancer as “any genital use,” a possible negative control for genital application and the hypothesized vaginal-uterine-fallopian retrograde migration hypothesis. The IARC publication [4] (at p.97) noted: “A key assumption of the use of negative control exposures is that any tendency for reduced or exaggerated recall of exposure is likely to be similar for the main study exposure and the negative control exposure.” Furthermore, Schildkraut et al. [36] also stratified results based on the year in which information on historical talcum powder use was obtained. Among those interviewed after 2014 (a date by which the authors presumed that publicity surrounding several ovarian cancer and talc litigation cases in the US would be widely known), the OR associated with “any genital use” increased from 1.44 (95% CI 1.11–1.86) to 2.91 (95% CI 1.70–4.97), whereas for “only nongenital use” it decreased from 1.40 (95% CI 0.96–2.03) to 1.26 (95% CI 0.69–2.32). They noted that they were unable to derive OR estimates for the subgroup “only genital use” “due to sample size considerations” – presumably too few women reported only genital talcum powder use. This example, consistent with the authors’ interpretation, provides support for the role of recall and reporting bias not only in general but also operating differently over time and possibly in response to publicity.

Because the bias analyses reported by the IARC monograph [16] and the previous one by Goodman et al. [32] based only on the Cramer et al. [37] case-control study used SQBA methods – neither of which could appropriately characterize bias parameter uncertainties – we chose to conduct a BQBA based on the pooled data from all of the pooled population-based case-control studies. Indeed, in addition to these improvements, IARC defined two advantages for which the Bayesian method would be preferred (see Table 4.8) [4]: (a) integration of prior knowledge about model parameters other than bias parameters and (b) flexible specification of the model beyond standard choices. Both are indicated with respect to the population-based studies on ovarian cancer and genital talcum powder use.

We excluded the published case-control studies nested in cohort studies because the exposure was estimated prior to any ovarian cancer (or other disease) diagnosis. We also excluded the hospital-based case-control studies. This may be especially important where, for example, people who are gravely ill may tend to make a greater effort in recalling and reporting exposures and assign greater importance to a perceived encounter with a presumed toxic agent compared with those in comparatively good health or unaware that an exposure may be hazardous and therefore forget or discount such encounters. In the meta-analysis reported by Kadry Taher et al., [20] data from four hospital-based case-control studies combined [38–41] did not find an increased meta-OR for genital talcum powder use (OR 0.96, 95% CI 0.78–1.17). The hospital-based case-control studies possibly provide a methodological refinement to reduce recall bias as the controls were women with newly diagnosed non-gynecologic malignancies [41] or with other non-gynecologic or malignant diseases [39].

In finalizing our decision to focus on the population-based case-control studies, we discovered that the crude OR for “ever use” of genital talcum powder as reported by Davis et al. [23] was 0.96 (95% CI: 0.88–1.05), consistent with the reported prevalence of genital talcum powder use of 31.4% in controls and 30.6% in cases (see Table 2 in Davis et al. [23]). However, the reported multivariate-adjusted OR was 1.32 (95% CI: 1.17–1.48) and the striking difference between these was not explained, perhaps reflecting some additional role of confounding. Further assessment of confounding preferably using BQBA seems warranted given the IARC monograph [16] only conducted a SQBA for confounding, and the Risk Assessment Committee (RAC) [35] of the European Chemicals Agency (ECHA) conducted none.

The strengths of our evaluation primarily include the large sample size, and the superiority of the Bayesian method for this application. Weaknesses include the inherent limitations of the exposure information derived from each study, not only due to the reliance on historical recall and reporting by cases and controls, but also the limited details that were obtained beyond “ever” versus “never” use in some studies, and the basic lack of any gold standard for any of the exposure information. Our BQBA may be vulnerable to alternative assumptions about the distribution of misclassification parameters, despite being robust to gross misspecifications. However, it must be noted that we did not elicit priors on misclassification parameters but used the result of such an effort by IARC [16]. If alternative defensible priors can be independently articulated, this would warrant future sensitivity analyses to the postulated priors, easily implemented using our explicitly documented methodology. We assumed homogeneous misclassification parameters across studies, as the data were too limited to support reliable stratified or heterogeneous bias-adjustment models by study period or publicity era. This assumption may be relaxed in future work with more detailed validation data. Also note that assessment of non-differential exposure and confounding biases was beyond the scope of our study.

Nevertheless, our results clearly demonstrate that reliance on the reported results of population-based case-control studies of ovarian cancer and genital talcum powder use for public health or regulatory or decision-making purposes is not scientifically justified and may be misleading. Any use of such studies to quantify risk and entertain causal inference, should be preceded by state-of-the-art quantitative bias analysis. While other potential sources of bias may be operating within this body of studies – and should be evaluated – we have demonstrated that reporting bias likely was the primary reason for the statistically significant ORs reported in the population-based case-control studies.

Authors’ contributions

All authors contributed to writing the main text. All authors have approved the submitted version. Each author agreed both to be personally accountable for the author’s own contributions and to ensure that questions related to the accuracy or integrity of any part of the work, even ones in which the author was not personally involved, are appropriately investigated, resolved, and the resolution documented in the literature. D.W. and I.B. prepared figures. W.J.T., K.A.M., and I.B. extracted data from the literature. Statistical analyses were carried out by D.W., W.J.T., J.Q., and I.B.

Funding

This work was financially supported by EUROTALC, a trade association. According to their website EUROTALC is “the talc industry’s representative body for all regulatory and scientific matters” and is a “member of IMA-Europe, the European industrial minerals association” (https://eurotalc.eu/). Neither EUROTALC nor any of its officers or members viewed this paper or any parts prior to submission.

Data availability

This manuscript uses publicly available summary data from the identified epidemiological studies. The computer code used for the analyses is also provided.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

K.A.M. and W.J.T. have served as expert witnesses for the Defense in litigation involving talc and cancers including ovarian, and K.M. provides epidemiological consulting to EUROTALC, a trade association of talc producers and users and the financial sponsor of this work. D.W., I.B. and J.Q. declare no potential conflicts of interest.