Sociodemographic disparities and potential biases in persistent pain estimates: Findings from 5 waves of the Irish Longitudinal Study on Ageing (TILDA)

Pain is a prevalent, debilitating condition among older adults. Much evidence on this topic comes from cohort studies, which may be affected by attrition and measurement bias. Little is known about the impact of these biases on pain estimates for European older adults. Additionally, there is a lack of longitudinal research on pain and sociodemographic disparities in Irish older adults.


| INTRODUCTION
Pain is estimated to affect 30%-60% of adults aged ≥50 years in Europe (Zimmer et al., 2020).Negative effects of pain include isolation, disability, and reduced quality of life (Breivik et al., 2013;Cohen et al., 2021).Internationally, studies of "older adults" (defined here as those aged 50 and older) typically find that the burden of pain is greater for certain sociodemographic subgroups, such as women and socioeconomically disadvantaged groups (e.g., those with lower education attainment) (Cimas et al., 2018;Stewart Williams et al., 2015).
Much evidence about pain comes from population cohort studies.However, these can be subject to attrition biases (Biele et al., 2019;Metten et al., 2022) if participants with certain characteristics are more likely to be lost to follow-up or die, which may affect estimates of pain prevalence, severity, or disparities.For example, large-scale volunteer databases of ageing are less likely to retain less healthy, less socioeconomically advantaged participants (Brayne & Moffitt, 2022).Mortality bias arises if excess attrition of certain sociodemographic subgroups occurs specifically due to death.
Grol-Prokopczyk (2017) explored the potential impact of bias due to mortality or non-response when estimating sociodemographic disparities in pain for older American adults, using multi-wave data.While no evidence of association between pain and non-response was found, pain severity was strongly predictive of mortality, suggesting that pain estimates may be subject to mortality bias.A similar association between severe pain and increased mortality risk was found in a Scottish older adult cohort (Torrance et al., 2010).
Measurement bias arises from systematic differences in self-reporting styles (reporting heterogeneity).There is a growing literature examining sociodemographic differences in reporting styles for subjectively rated health conditions, including pain (Bago d'Uva et al., 2008;Chan et al., 2011;Ziebarth, 2010).Failure to recognize differences in reporting styles could impact the validity of relative rankings and comparisons between groups (Menec et al., 2007).Estimating the presence and direction of such self-reporting bias is thus an important task, sometimes undertaken by comparing self-reports to more objectively recorded measures of condition (Jürges, 2007;Spitzer & Weber, 2019).
Limited research has investigated the impact of biases due to attrition and specifically mortality (Lacey, Jordan, et al., 2013;Muszyńska-Spielauer & Spielauer, 2022) and reporting heterogeneity (Spitzer & Weber, 2019) on health status estimates among European adults.To our knowledge, the impact of these biases on estimates of sociodemographic disparities in pain has not been explored in any European cohorts.Research on the demography of pain in Ireland in particular is sparse, consisting of a small number of cross-sectional and longitudinal analyses examining associations between older adult pain and factors including health and healthcare utilization (Kennedy et al., 2017;O'Neill et al., 2018O'Neill et al., , 2020;;Raftery et al., 2011).
Establishing accurate, unbiased estimates of pain prevalence and sociodemographic disparities is crucial to inform public policies targeting this potentially pervasive and debilitating condition.This study uses five waves of the Irish Longitudinal Study on Ageing (TILDA) to contribute to this goal by (1) examining previously unexplored longitudinal sociodemographic disparities in pain among older Irish adults and (2) investigating how attrition bias, mortality bias, and reporting heterogeneity measurement bias may affect the accuracy of such pain estimates.

| Population and participants
This study is a secondary analysis of five consecutive waves of TILDA.TILDA is a nationally representative cohort study of the health, social, and economic conditions of community-dwelling older adults in the Republic of Ireland.A multi-stage sampling design was used to select the baseline (Wave 1) sample.The first step involved grouping all residential addresses in the Republic of Ireland into 3155 townland clusters, using the Irish Geodirectory as a sampling frame.These clusters were stratified by socioeconomic group and geography, and a representative sample of 640 clusters was selected.A probability sample of 40 addresses was then drawn from each cluster, and contact was made to recruit household members aged 50 and over.A response rate of 62% was achieved at the household level, with 8171 community-dwelling older adults and 329 of their younger partners participating in the study at Wave 1 made available for analysis.Wave 1 commenced in 2010 and subsequent waves of data collection occurred every 2 years.At each wave, participants were invited to complete a computer-assisted personal interview, educational attainment, and those without private health insurance were found to have the highest pain burden longitudinally, suggesting a need for targeted interventions for these groups in Ireland and internationally.a self-completion questionnaire, and a health assessment.Health assessments were carried out in one of two TILDA health centres or in the participant's home.The full study design is detailed elsewhere (Kenny et al., 2010).
Individuals who were aged 50 or over at Wave 1 of TILDA were included in the longitudinal analysis.Younger partners were excluded.Of the 8171 participants in our Wave 1 sample, 6993 (85.6%) returned for Wave 2; 6246 (76.4%) for Wave 3; 5571 (68.2%) for Wave 4; and 4872 (59.6%) for Wave 5.This closed cohort design ensured that longitudinal data across the five waves would be present for all participants, except in the case of non-response or death.We defined attrition as participants leaving the cohort between waves, either due to death or other loss to follow-up.

| Pain
At each TILDA wave, participants were asked, "Are you often troubled with pain?" (yes, no).We label this pain phenotype "persistent troubling pain" ("pain" for brevity).While a duration of pain is not specified, previous research suggests this question is unlikely to capture acute or transient pain.One study found respondents were more than twice as likely to report "any pain in the last 30 days" as to report being "often troubled by pain" (Banks et al., 2009).Those who answered yes to this initial pain question were then asked, "How bad is the pain most of the time?" (mild, moderate, severe).Responses to both pain questions were combined to make a 4-category "pain status" variable for each wave.For some parts of the analysis, these pain status categories were converted to a numerical pain score using the following codes: 0 = no pain, 1 = mild pain, 2 = moderate pain, 3 = severe pain, as in previous studies (Dunn et al., 2006;Grol-Prokopczyk, 2017), and scores averaged across groups.Those who reported being often troubled by pain were also asked "Does the pain make it difficult for you to do your usual activities such as household chores or work?" (yes, no).This question was used as an indicator of experiencing pain-related disability or not.These pain questions have been used previously as measures of experiencing pain, pain severity, and pain disability respectively in older Irish adults (O'Neill et al., 2020) and in older adult populations worldwide (Bell et al., 2022;Mohanty et al., 2022).
This study focused on non-cancer pain only.Responses to pain variables were set to missing in cases where reported pain was likely due to cancer or cancer treatment.This rule affected between 0.1% and 1.2% of the sample at each wave.

| Sociodemographic factors
Demographic variables were age category (50-59, 60-69, 70-79, 80+), sex (male, female), highest level of education (none or primary, secondary, tertiary), and whether individuals were covered by a private health insurance policy (yes, no) at baseline.Education level and private health insurance status were used as proxies for socioeconomic status (SES); while TILDA does collect data on income and assets, these are answered only by a subset of the sample; education and insurance are answered by all participants and therefore by using them we minimize missing data issues.Irish healthcare services are financed both publicly and privately, with the primary benefit of private health insurance being reduced wait times for elective hospital treatments (Turner & Smith, 2020).47% of the Irish population had private health insurance at the end of 2021 (Health Insurance Authority, 2023).

| Attrition by wave 5
Attrition by Wave 5 was defined as someone not participating in Wave 5 for any reason, including death.As part of sample maintenance efforts, the TILDA team attempted to contact and invite baseline participants for interview at each follow-up wave even if they had missed a previous wave, unless the participant had been confirmed deceased or requested to withdraw from the study (Donoghue et al., 2017).If someone missed an intermediate wave/ waves after Wave 1 but had returned to the study by Wave 5, they were included as present in the Wave 5 sample.The number of Wave 5 participants who had missed at least one prior wave was small (n = 278, 5.7% of those present at Wave 5).

| Mortality data
Mortality data included survival status (confirmed deceased or not confirmed deceased) and year of death if confirmed deceased.To obtain this data, the TILDA team performed linkage between General Register Office death records and individual-level survey data.The full linkage process is detailed elsewhere (Ward et al., 2020).The linkage identified 741 deaths in our sample between the end of Wave 1 and March 2018 (Wave 5 data collection commenced in January 2018).By comparing year of death to year of data collection for each wave after Wave 1, a survival status variable was created for each follow-up Wave (2-5), indicating whether a participant was still alive at that wave.When a participant died between waves, their survival status was set to alive at the prior wave and deceased at all following waves.For example, a participant who died in 2013 was coded as alive at Wave 2 (collected in 2012) and deceased from Wave 3 (collected in 2014/2015) onwards.

| Statistical analysis
2.3.1 | Descriptive statistics Unweighted descriptive statistics were reported as counts and percentages for categorical sociodemographic, pain status, and survival status variables at each wave.Note, some weighted pain prevalence statistics are available in the TILDA literature (Barrett et al., 2011).
2.3.2 | Biases due to attrition and mortality An alluvial plot was used to visualize transitions in pain status (none, mild, moderate, severe), survival status, and lost to follow-up categories across the five waves.
Next, to investigate attrition bias, we fitted a logistic regression model of attrition (due to death or otherwise lost to follow-up) on pain severity.Attrition by Wave 5 was used as the outcome, with pain severity at Wave 1 as a predictor and controlling for Wave 1 sociodemographic factors.To investigate attrition bias due to mortality specifically, we fitted a logistic regression model of mortality by Wave 5 on pain severity, controlling for baseline sociodemographic factors.Those who were lost to follow-up by Wave 5 but not confirmed deceased were removed from this analysis.Adjusted odds ratios (AORs) with 95% confidence intervals (CIs) are reported for all logistic regression models.The likelihood ratio chi-squared test statistic and McFadden's pseudo R-squared (McFadden, 1979) are reported as measures of model fit.

| Reporting heterogeneity
Reporting heterogeneity is a form of measurement bias due to systematic differences in how distinct population subgroups self-report subjective conditions-in the case of our study, pain (Molina, 2016).Using logistic regression and Wave 1 data, we estimated the presence of reporting heterogeneity by examining differences across sociodemographic groups in the odds of reporting painrelated disability for a given severity of pain.Groups more likely to report pain-related disability at a given pain level may be experiencing pain that has higher impact and thus may be "stoical" in their expression of pain severity.An AOR above 1 for a given sociodemographic group suggests such stoicism, as members of this group are more likely to report pain-related disability for a given level of pain, compared to the reference group.This method of assessing differences in pain reporting styles is modelled on the approach in Grol-Prokopczyk ( 2017), but is exploratory rather than definitive; limitations of and alternatives to this approach are acknowledged in the Discussion.While not perfectly correlated, selfreported pain severity is strongly associated with painrelated disability in both clinical and population-based studies.A strong dose-response relationship has been observed whereby greater pain severity is predictive of greater disability (Covinsky et al., 2009), including for objective measures of function limitations such as gait speed (Simmonds et al., 2012).

| Longitudinal sociodemographic disparities in pain
To visually examine sociodemographic disparities in the reporting of pain, pain scores were calculated on the 0 to 3 scale, where 0 = no pain and 3 = severe pain.Differences in average pain scores across sociodemographic factors were then summarized at each wave using stratified line plots.Similarly, a line plot of average pain scores for survivors, decedents, those lost to follow-up, and the full sample across all waves illustrated if decedents had higher average pain scores than survivors.A multivariate latent growth curve model with pain severity status as an ordinal response (ordered categories: no pain, mild pain, moderate pain, severe pain) was fitted to estimate sociodemographic differences in pain trajectories.In brief, the model estimates the continuous latent response variable assumed to underlie the observed ordinal response variable.The latent pain trajectory is estimated in terms of an intercept and slope.Sociodemographic differences in pain trajectory are modelled as deviations from this intercept and slope.Thresholds are estimated to delineate the ordinal pain categories on the continuous latent variable scale.We assumed longitudinal threshold invariance, meaning the thresholds for determining pain severity category from the latent variable were the same at each wave (Masyn et al., 2014).Root Mean Squared Error of Association (RMSEA) and Comparative Fit Index (CFI) are reported as measures of model fit.As a rule of thumb, RMSEA values below 0.08 and CFI values above 0.95 are considered indicative of "good fit" (Hooper et al., 2008).

| Descriptive statistics
Table 1 details the number (%) of those remaining in the study at each follow-up wave across sociodemographic factors measured at Wave 1 and the recorded pain severity and pain-related disability across the waves.At Wave 1, 60% of participants were aged 60 or over, with 7.7% of the sample aged 80 or over.Just over half (54.2%) of the participants were women.Secondary was the most common highest level of education (39.9%), followed by none or primary (30.6%) and tertiary (29.4%).A majority (57.5%) reported having private health insurance.
By Wave 5, the proportion of the remaining sample who had reported private health insurance at Wave 1 was 64.2%.Of the Wave 5 sample 55.4% were aged 60 or over at Wave 1, while those aged 80 or over at Wave 1 represented just 3.1% of the Wave 5 sample.Secondary remained the most common highest Wave 1 education level by Wave 5 (41.6%), but tertiary became more common than none or primary education level (35.1% and 23.3% respectively).Just over half (55%) of the Wave 5 sample were women.
At Wave 1, 35.3% of participants reported they were often troubled by some degree of pain.The most commonly reported pain severity was moderate pain (16.5% of the full sample), followed by mild (10.1%) and then severe pain (8.7%).This distribution of pain severity remained very consistent across all five waves.This is evident in the latent growth curve model in Table 6 (discussed in more detail later), where the slope/rate of change in pain over time is not found to be significant.However, this finding may be affected by attrition biases.These potential sources of bias are explored further in the following sections.The percentage of those with pain who reported pain-related disability was around 60% across waves.For those participating in any given wave, missing data for pain questions were low with item non-response at most 2.3%.Table 2 summarizes the percentage attrition at each wave, with a breakdown of what percentage of participants were confirmed deceased or otherwise lost to follow-up.14.4% of the baseline sample did not return for Wave 2. This percentage increased to 23.6% at Wave 3 and 31.8% at Wave 4. By the beginning of Wave 5, a large proportion (3299, 40.4%) of baseline participants had left the study, either due to death or other loss to follow-up.741 (9.1%) members of the baseline sample were confirmed deceased by this time, which accounted for 22.5% of the missing cases.

| Attrition and mortality bias
Figure 1 is an alluvial plot showing the percentage in each pain category and transitions between pain categories across the five waves.Lost to follow-up (not confirmed deceased) and confirmed deceased categories are also included to capture sample attrition.The plot highlights the extent of mortality and lost to follow-up between waves.Of the remaining sample at each wave, the ratio of no pain to mild to moderate to severe pain appears consistent, as seen in Table 1.
The alluvial plot also visualizes pain category transitions over time.While there is considerable transition between categories, the majority of individuals at a particular wave tend to either stay in the same pain category or move to an adjacent pain category in the subsequent wave, excluding those who dropped out or died.For example, if we observe the severe pain category, the most common transitions between waves were to the same category or to the moderate pain category.
Results of the multivariable logistic regression of attrition (due to death or loss to follow-up) are presented in Table 3. Pain severity at Wave 1 was not a significant predictor of attrition by Wave 5, controlling for sociodemographic factors.However, those with a tertiary-level education at Wave 1 had considerably lower odds [AOR 0.48 (95% CI 0.42, 0.55)] of exiting the study by Wave 5 than those with none or primary education level, controlling for other factors including age.Those with secondary-level education [AOR 0.67 (95% CI 0.59, 0.75)] also had lower Table 4 presents the results of the multivariable logistic regression of death by Wave 5, comparing those who were confirmed deceased with those still participating at Wave 5, using pain severity at Wave 1 as a predictor while controlling for sociodemographic variables.Severe pain at Wave 1 was associated with a 63% increase in odds of death by Wave 5 [AOR 1.63 (95% CI 1.20, 2.19)] compared to those who were pain-free at Wave 1, controlling for age, sex, education level and health insurance status.Those with a secondary [AOR 0.70 (95% CI 0.56, 0.87)] or tertiary [AOR 0.63 (95% CI 0.48, 0.81)] level education at Wave 1 had lower odds of dying by Wave 5 than those with none or primary education level, while those without private health insurance had an estimated 75% higher odds of dying [AOR 1.75 (95% CI 1.43, 2.13)] compared to those with private health insurance.

| Reporting heterogeneity measurement bias
Differences in reporting pain-related disability between sociodemographic groups for given levels of pain severity at Wave 1 were examined as evidence of the presence of reporting heterogeneity using logistic regression, as given in Table 5.The analysis found the AOR for reporting pain-related disability of those without private health insurance cover was 1.37 (95% CI = [1.15,1.64]), suggesting that these individuals were more impacted by experience of pain and may be more stoical in their expression of pain severity, compared to those with health insurance.Other sociodemographic characteristics were not significantly predictive of pain reporting style.

| Longitudinal sociodemographic disparities in pain
Figure 2 shows average pain scores in each wave stratified by Wave 1 sex, age category, highest education level, and health insurance status, and by survival status by Wave 5.In general, the trend lines are roughly flat, suggesting that average pain severity neither increases nor decreases over time within the different groups.However, average pain scores appear to vary considerably across the different sociodemographic groups.

AOR (95% CI)
Baseline sociodemographic characteristics 2558 Wave 1 participants who were lost to follow-up by Wave 5 but not confirmed deceased were removed.
Average pain were consistently higher for women men at all waves of TILDA (Figure 2a).The average score for women was approximately 0.80 (on a scale from 0 to 3) in all waves, versus 0.55 for men.The differences in average pain scores between Wave 1 age categories across waves, Figure 2b, were small compared to the differences observed for sex.The youngest group (aged 50-59 at baseline) had the lowest average pain scores at approximately 0.65, followed by the 60-69 and then the 70-79 age groups.The average pain scores for the oldest group (aged 80+ at baseline) varied considerably, perhaps due to mortality between subsequent waves decreasing the size of the group.Average pain scores for those aged 80+ at Wave 1 were the highest out of all the Wave 1 age categories, peaking at approximately 0.75.
There were clear and consistent disparities in pain scores across educational groups (Figure 2c).Those with no or primary levels of education at baseline consistently reported experiencing higher levels of pain, with average scores between approximately 0.75-0.85across waves.Each subsequent increase in education level corresponded to lower average pain scores, approximately 0.65 for those with secondary level education and 0.55 for tertiary level education.Figure 2d, which depicts average pain scores for those with and without private health insurance at baseline, reveals similar disparities in pain over time.The average pain score for those without private health insurance (approximately 0.85) is higher than for those with it (approximately 0.60) at all waves.
Results from the multivariate latent growth curve model of longitudinal sociodemographic disparities in pain are shown in Table 6.RMSEA = 0.015 and CFI = 0.998 indicate very good model fit.The estimated thresholds for the ordinal categories mild pain, moderate pain, and severe pain are 0.572, 0.870, and 1.649 respectively.The model parameters reflect the trends observed in the Figure 2 plots, including the relative flatness of the trend lines.All slope terms are not significantly different from zero, suggesting that propensity to experience pain is neither increasing nor decreasing overall or within the different sociodemographic groups over time.
However, intercepts did differ significantly by sex, education, and health insurance type.Women had a higher baseline propensity for pain [intercept 0.282 (95% CI 0.213, 0.351)] than men.Compared to those with no or primary level education, propensity for pain was successively lower for those with secondary Figure 2e shows average pain scores for those who died during the 8-year Wave 1 to Wave 5 study period (decedents), as well as the average pain scores for those who left the study before Wave 5 but were not confirmed deceased (lost to follow-up) and those who participated until Wave 5 (survivors).Average pain scores were highest for decedents at baseline then followed a negative linear trend, dropping below the average for those lost to follow-up or survivors by Wave 4. This suggests that those who had more severe pain at Wave 1 were more likely to die earlier in the study period and thus leave the study, inducing bias due to mortality.Survivors had the lowest average pain scores of the three groups up to Wave 4, and were consistently lower than the lost to follow-up group.
Figure 3 stratifies the decedent group by time period of death.In general, the shorter the period of survival, T A B L E 5 Multivariable logistic regression of pain-related disability on sociodemographic factors, controlling for pain severity, using Wave 1 data (n = 2885) a .

AOR (95% CI)
Wave 1 sociodemographic characteristics the higher average pain score Wave 1 (with the exception of the "Died Wave 2" group, whose baseline average pain score was very close to the "Died Wave 4-Wave 5" group).We also note that average pain scores were consistently higher for those who were lost to follow-up than those who continued to participant to Wave 5.As participants who died or were otherwise lost to follow-up had higher mean pain scores than those who participated through Wave 5, those with higher pain levels become underrepresented in the sample over time.These findings suggest evidence of bias due to attrition.

| DISCUSSION AND CONCLUSIONS
Pain in older adults is a pervasive public health problem (Cohen et al., 2021) associated with negative outcomes including functional disability (Makino et al., 2019), frailty (Wade et al., 2017), reduced quality of life (Ludwig et al., 2018) and increased mortality risk (Torrance et al., 2010).Deriving accurate estimates of older adult pain prevalence and severity for different sociodemographic groups is important to guide policymakers across Ireland and Europe in this problem.Our suggests that sex and socioeconomic disparities in pain found in other countries (Cimas et al., 2018;Grol-Prokopczyk, 2017;Ikeda et al., 2019;Lacey, Belcher, et al., 2013;Stewart Williams et al., 2015) are mirrored in Irish older adults.We also found evidence that Irish older adult pain estimates are subject to biases including reporting heterogeneity and mortality bias.

| Attrition and mortality bias
We found significant risk of pain-related mortality bias in TILDA, consistent with other longitudinal ageing studies.As those with more severe pain were more likely to die earlier in the study, there is a disproportionate loss of participants with severe pain.This mortality bias has serious implications.For example, studies of what factors predict will biased if the data excludes people who already died and people had higherthan-average pain.Additionally, this bias would lead to underestimation of the increase in pain with age in cross-sectional data, as people with more pain are more likely to die and so it will appear that pain risk does not rise with age.Pain severity was also a significant predictor of death in American (Grol-Prokopczyk, 2017) and Scottish older adult cohorts (Torrance et al., 2010).This combined evidence of pain-related mortality bias in older adult cohorts from different continents has important implications for pain research, requiring researchers to identify and mitigate such bias to avoid underestimating older adult pain experiences over time.Interestingly, while only severe pain was predictive of death in TILDA and the Scottish study, both moderate and severe pain were significant mortality predictors in the American analysis.This may suggest a stronger association between pain and death among American older adults.
We did not find statistically significant evidence of attrition bias in the TILDA cohort.However, while the effect estimates of pain severity on attrition were not large in either TILDA or the HRS (Grol-Prokopczyk, 2017), painrelated attrition bias should not necessarily be ruled out as a potential source of bias in older adult pain studies.A feature common to both TILDA and the HRS is that the samples consisted entirely of community-dwelling older adults at baseline, but participants who were later institutionalized were interviewed at subsequent waves when possible.As suggested in Grol-Prokopczyk (2017), studies that instead remove participants who move to care institutions from the follow-up sample may induce attrition bias, as participants with more severe pain may be lost.If present, attrition bias would result in the underestimation of pain prevalence.
Lower educational attainment and not having private health insurance were associated with increased odds of both mortality and attrition in TILDA, after controlling for age.These socioeconomic subgroups become underrepresented at later waves, an issue common in large voluntary studies of ageing (Brayne & Moffitt, 2022) which may induce bias.These findings highlight a need for initiatives to tackle socioeconomic barriers to sample recruitment and retention to reduce the risk of bias.Potential strategies include oversampling groups whose members are more likely to drop out, offering compensation or free travel to complete interviews and attend health assessments, and offering translated questionnaires for non-English speaking participants (Bonevski et al., 2014).
As some degree of attrition is typically unavoidable, we also highlight a need for awareness around the potential for such biases and methods to mitigate them.Multiple imputation (MI, Rubin, 1987) is one popular approach for handling missing data.However, standard MI methods assume the data is missing at random (MAR), meaning the missingness depends only on the observed data (Van Buuren, 2018).As we found evidence that more severe pain is associated with mortality (missingness), it is plausible that the missingness is dependent on the missing (unobserved) pain values themselves.In this case, the data is missing not at random (MNAR), and results can be sensitive to violation of the MAR assumption (Carreras et al., 2021).Causal analysis designs such as instrumental variables (Tchetgen Tchetgen & Wirth, 2017) and inverse probability of censoring weighting (Rotnitzky et al., 1998) may be alternatives to handle this MNAR data.However, future research would be needed to examine the potential for these approaches to address the bias issues raised in this study.

| Reporting heterogeneity
We also found evidence of reporting heterogeneity, or measurement bias due to differences in self-reporting styles, at TILDA Wave 1.We interpret the results as those without private health insurance being more stoical in their reporting of pain than those with it.This result was similar to the multi-wave pooled analysis in Grol-Prokopczyk (2017), where participants with lower SES were found to be more stoical.While potential associations between lower SES and "good patient" behaviour have been posed previously (Pillay et al., 2014), there has been little work explicitly examining the extent and direction of reporting heterogeneity of persistent pain across sociodemographic groups.An Austrian study found a similar association between lower SES and disability while controlling for pain level, though the finding was labelled "unexplained" rather than attributed to possible reporting heterogeneity (Dorner et al., 2011).These findings may suggest a tendency in both Europe and America for socioeconomic pain disparities to be underestimated, due to those in less advantaged circumstances being stoical in their pain reporting.However, unlike in the HRS study, women and those with no or only primary education were not found to be significantly more stoical in our TILDA analysis.This may suggest that the reporting habits of older adults and the propensity to be more stoical may vary by culture and geographical location.Determining the extent and direction of cross-group reporting heterogeneity should be a key consideration for any population study interested in group comparison, although strategies to overcome this type of bias, such as anchoring vignettes, are not straightforward (Grol-Prokopczyk et al., 2015).

| sociodemographic disparities in pain
Our using Irish data confirm from other countries that the burden of pain in older adults is worse for women and those with lower SES (Jacobs et al., 2006;Milani et al., 2022;Palacios-Ceña et al., 2015;Wranker et al., 2016).Average pain scores and propensity for pain were consistently higher for women compared to men across waves.Those without private health insurance and those with lower levels of education also had consistently higher average pain scores and propensity for pain than their higher SES counterparts across all waves.It is possible that age confounded some of the disadvantage for participants with a lower level of education in the trend plot, whereby those who were older appeared more likely to have a lower level of education and a higher risk of mortality.However, differences in propensity for pain across the education and health insurance categories remained significant after controlling for age in the multivariate latent growth curve model, while we found no significant evidence of age-related pain disparities in Irish older adults.Some international trends were not reflected in the TILDA data.Research on other European countries has found that most (but not all) countries show a positive increase in pain prevalence over time, net of age (Zimmer et al., 2020).Similar increasing pain trajectories over time have been reported in the US (Grol-Prokopczyk, 2017) and Canada (Shupler et al., 2019).In contrast, pain scores remained relatively flat across waves in TILDA and rates of change in propensity for pain were not significant.These differences may reflect genuinely different pain patterns among Irish older adults, perhaps due to differences in lifestyles and healthcare provision, for example.Alternatively, the flattened pain trend lines for TILDA may reflect the mortality bias suggested in our results.
Due to the relative homogeneity of the current older Irish population, we did not explore potential racial or ethnic pain disparities.However, we note that racial/ethnic disparities have been found in countries with more diverse older populations, such as the US (Morales & Yong, 2021).Such disparities may become relevant to Ireland's policymakers as the diversity of the population increases.
A limitation of our study was the lack of an automated linkage system between survey data and death registration in Ireland.There is a time lag in the registration of deaths, so some participants who died may be categorized as "lost to follow-up" rather than "deceased" in our analyses.Without an objective measure of pain severity, evidence for the presence of reporting heterogeneity was sought as differences in pain-related disability experienced by groups with the same level of reported pain severity.This assumption that pain-related disability was reported without bias and only pain severity was subject to reporting heterogeneity is another limitation.The results of our reporting heterogeneity analysis are therefore exploratory.Future work is required to validate these findings and further examine differences in pain reporting styles.This could be done by comparing pain self-reports to objectively timed walking speeds, which are highly correlated with pain severity (Hicks et al., 2017;Simmonds et al., 2012), or to measures such as frailty indices, which have been used to assess discrepancies in self-rated health (Calvey et al., 2022).The measures of pain prevalence ("are you often troubled by pain?") and pain severity ("How bad is the pain most of the time?") are also somewhat limited, as is the dichotomous indicator of pain-related disability, which does not convey degree of disability.However, these pain questions are used across multiple global ageing cohort studies, which has the benefit of allowing direct replication and comparison of our results across ageing populations in different countries (Gateway to Global Aging Data, 2023).Finally, reporting heterogeneity was explored using Wave 1 data only.Future research could investigate changes in reporting heterogeneity over time.
A key message from this work is that the potential for bias in population studies cannot be ignored.Failure to investigate and account for such biases may result in inaccurate estimates of pain prevalence and pain disparities in older adult populations, weakening the evidence base which guides policymakers' decisions.Identifying bias is also an important step for potential future work looking at causal relationships between pain and attrition.Additionally, we highlight a need to address sociodemographic disparities in pain among Irish older adults.Targeted interventions are required to tackle the disproportionate pain burden of women, those with lower levels of education and those without private health insurance.
Many countries across Europe and the world have ageing populations, which present economic and healthcare challenges (Christensen et al., 2009).Accurate estimates of pain prevalence, along with an understanding of pain trajectories and sociodemographic disparities in pain, will be required for policymakers and health services to plan appropriately.This work is a first step towards providing such estimates for the older Irish population, while highlighting biases that may impact pain research using observational studies in Ireland and internationally.

AUTHOR CONTRIBUTIONS
ER, AH, and HP conceptualized the design of the study.ER conducted the analyses, interpreted the results, and wrote the first draft of the manuscript.PM supported access to the data.All authors interpreted the results, Summary of attrition (deceased or otherwise lost to follow-up) at each wave.Transitions in pain status categories, including deceased and lost to follow-up, between waves.The thickness of the streams between pairs of categories at adjacent waves is proportional to the number of participants who transitioned between those two categories at those waves.odds of Not having private health insurance at Wave 1 was associated with increased odds of attrition by Wave 5[AOR 1.45 (95% CI 1.32, 1.61)].
Multivariable logistic regression of attrition (due to death or otherwise lost to follow-up) by Wave 5 on pain severity and sociodemographic factors at Wave 1 (n = 8150) a .
[intercept −0.126 (95% CI -0.216, −0.036)] and tertiary [intercept −0.228 (95% CI -0.326, −0.130)] level of education.Propensity for pain was also significantly lower for those with private health insurance at Wave 1 [intercept −0.239 (95% CI -0.313, −0.165)] than those without.Only the intercept terms for age categories were not significantly different to zero, suggesting no age disparities in pain.The variance of the latent variable intercept was significantly different to zero [0.708 (95% CI 0.679, 0.737)], indicating significant individual variation across participants in their propensity for pain.The variance in slope/rate of change in propensity for pain across participants was also significant [0.024 (95% CI 0.020, 0.028)].The covariance between the latent variable intercept and slope suggested a weak inverse relationship between Wave 1 propensity for pain and rate of change over time [−0.045(95% CI -0.055, −0.035)].

F I G U R E 2
Average pain score over time by sociodemographic factors.
Abbreviations: CFI, comparative fit index; RMSEA, root mean squared error of association.
Descriptive statistics for Wave 1 sociodemographic factors and pain severity across waves in TILDA Waves 1-5.
T A B L E 1a Percentages are for the number of participants who reported being often troubled by pain at each wave, not the entire sample.