Using a large sample of U.S. older adults, we found no difference in the risk of Alzheimer’s disease associated with immune-mediated inflammatory diseases relative to the risk in a large sample drawn from the general US population. In subgroup analyses we found no increased risk of AD associated with RA, PSA, and AS; however, we recommend cautious interpretation of disease-specific results due to small subgroup sizes. Our disease-specific models were unadjusted for important covariates due to these limitations. Our primary finding is that as a group, respondents with IMIDs had no increased risk of AD.
Many previous studies examined the risk of all-cause dementia and inflammatory autoimmune diseases, most commonly RA, whereas our study examined the risk of a specific type of dementia, Alzheimer’s disease, in a group of autoimmune diseases that included RA. Of prior studies that examined AD specifically, Wallin and colleagues found that when controlling for age, gender, and follow-up time there was an increased risk in RA; however, the association was at the borderline of statistical significance when additionally controlling for APOE-e4, smoking, and glucocorticoid/NSAID use [6]. Moreover, RA cases were identified via survey self-report, which we and several other studies have previously shown is an invalid RA measure [6, 39,40,41,42]. The use of self-reported RA classification casts doubt on the validity of these findings. Kao and colleagues found a statistically significant inverse association between RA and AD by retrospectively identifying a medical history of RA in current cases of AD [7]. This method may underestimate the RA-AD association because people with RA have an increased mortality risk and may not have survived to the point of AD measurement. Survivorship bias may explain why RA appeared to have a protective effect. Our study corrects for the shortcoming of both these studies by using a rigorous detection method for RA with two additional sensitivity measures, and by using time-to-event analysis which accounts for losses to follow-up due to death. In the case of IBD, Zhang and colleagues found an increases risk of AD in Crohn’s disease and UC, however, the study did not match participants on education and APOE-e4, two of the prominent risk factors of AD [11]. Similarly, Jang and colleagues found that AD risk is increased in AS, but likewise did not match on education or APOE-e4 [13].
In our analysis, age, education, and APOE-e4 were significant independent risk factors for AD, but not IMID’s as a group, or RA, PSA, and AS individually (with too few cases to estimate risk in Crohn’s and UC). Though in our sample, the frequency of race, education, and APOE-e4 did not differ between IMID and non-IMID groups, we chose to adjust for them based on the examination of the literature showing evidence of their associations with some IMIDs and cognitive impairment. Simulation studies show that when theoretical and empirical evidence differ for confounders in a dataset, that the theoretical confounders should still be adjusted for [43].
The strength of our study is the longitudinal design using a cohort of older adults from a nationally representative U.S. sample. An additional strength is the grouping of IMIDs that share a common inflammatory mechanism and common pharmaceutical treatments targeting this mechanism, which to our knowledge, has not been done before in research on cognitive impairment in groups of autoimmune or immune-mediated inflammatory diseases.
The most prominent limitation of our study is the use of claims-based algorithms to classify IMID and AD. The limitations of ICD-9 based algorithms come in two forms: error in coding sources and errors in validity. Coding errors may result from discrepancies between electronic and written records, miscommunication between patients and physicians, limitations in clinician’s knowledge of a specific illness, or unintentional recording errors [44]. Validity errors arise from whether or not diseases classified by an algorithm identify true cases of the disease, discussed further below.
Validity studies for the detection of specific autoimmune diseases in administrative databases are common. For example, prior research suggests that an algorithm of two ICD9-CM 714* diagnoses has a relatively low positive predictive value for detecting RA (PPV = the proportion of people identified by the algorithm with RA in the medical record) [45]. However, our study includes all of the diseases falling under 714*, not just RA. Therefore, our algorithms classify someone with RA or a related condition with any two of the 714* codes within 2 years, for example, two 714.9 codes for unspecified inflammatory polyarthropathy, or one code 714.0 for RA, and one 714.1 code for Felty’s syndrome. We did not allow counts across categories. The strength of our approach is that it includes several inflammatory autoimmune conditions in the IMID group. The limitations are that we do not know the validity of each disease-specific category. However, research shows that due to the difficulty in diagnosing many systemic autoimmune diseases, where false-positives occur in administrative data, those same subjects are often found to have a confirmed diagnosis of another related autoimmune disease [46]. Further, during the detection window in 2006–2009, the average number of claims-based diagnoses was 13.7 for RA, 7.3 for PSA, 6.4 for AS, 19.2 for Crohn’s disease, and 4.8 for UC, suggesting that those classified were receiving ongoing care. The prevalence of IMID in our sample (6.02%) is also within the estimated prevalence range of the general population [3].
In our sensitivity analysis, fewer people were detected as IMID and the HRs moved from above 1.0, to below 1.0; however, the results were statistically insignificant. The changing direction of the HRs suggests that calculation of AD risk in administrative databases is sensitive to IMID classification. Sensitivity to classification arises from validity tradeoffs depending on the strictness of a classification algorithm chosen. Less strict classifications may capture people with less severe disease who require infrequent care. At the same time, less strict criteria may include people who do not have an IMID (i.e., lower positive predictive value). More strict classification criteria in general will have higher positive predictive values, meaning a higher percentage of people detected truly have the disease, but lower sensitivity, meaning of those who truly have the disease, fewer are detected compared to less strict methods. A stricter criterion may increase the number of false negatives, moving some people who are correctly classified by a less strict algorithm to be incorrectly classified in the stricter algorithm, which may explain the changing direction of HRs in our analysis.
The same limitations present in the classification of IMIDs are present in the classification of AD. In addition to sources of coding error previously mentioned, prior research suggests that the validity of a single AD or related dementia hospital or physician-diagnosed code has 85.3% sensitivity, 94.2% specificity, 41% positive predictive value (PPV), and 99.3% negative predictive value (NPV) [47]. The relatively high sensitivity (85.3%) means most people who truly have AD will be identified as such, though the low PPV means that the algorithm identifies many people with AD who do not have it. Misclassifications of the outcome may explain our null findings compared to other studies using a different AD detection method.
Another potential limitation is the underdiagnosis of dementia in the population [48, 49]. The most prominent consideration for our results is whether under-diagnosis of dementia is differential or non-differential between our exposure and control groups. In the event there is differential classification, such that either the exposure or control group has a higher probability of under-diagnosis, than our results would be biased. We can only speculate here, however, those with IMIDs are likely in more frequent contact with clinicians than the general population, and therefore would likely encounter more opportunities for screening, referrals, and detection of dementia. If this does in fact occur, we would expect those with IMIDs to have a higher probability of dementia diagnosis, which would bias our results away from the null. We cannot say definitively how underdiagnosis differs between groups in our research and acknowledge this as a potential limitation of our results.
Another limitation is that the detection of both AD and IMIDs is limited to the timeframe in which beneficiaries have Medicare. The Medicare-restricted timeframe means some respondents could have had AD or IMID claims prior to receiving Medicare benefits, but not during our study. However, all the IMIDs in our study require ongoing care; therefore, it is unlikely someone would have an IMID prior to receiving Medicare benefits, but no indication of care thereafter. For AD, we began the study with a three-year detection window in which all respondents have complete FFS Medicare coverage meeting the CCW observation criteria for AD. We also required all respondents to have complete FFS coverage in years prior to 2006 in which they had Medicare benefits to exclude prevalent cases of AD. Because the majority of people with AD are 65 years of age and older and Medicare eligible [50], we do not believe this is a significant limitation likely to impact our results. A similar limitation is the 5-year AD detection window, which may be too short for the development of AD. However, the majority of our study population is 65 years of age or older, which has the highest incidence of AD, doubling every 5 years thereafter [51].
Residual confounding may also bias our results. Our models include covariates of age, age squared, gender, race, ethnicity, and APOE. Though we do not believe these covariates lack precision, it is possible our models do not include sufficient controls for other risk factors of AD that could affect our results. For instance, our models do not control for diabetes, cholesterol, smoking history, lifestyle factors, and other vascular risks that have been associated with AD [52]. Cardiovascular risk factors vary between different IMID’s, for instance, obesity is common in RA and PSA, but decreased in AS, UC, and Crohn’s disease [53]. Our models do not control for these possible confounders because we analyzed a group if IMIDs; accounting for variation in AD risk factors across the individual diseases of our grouping was not possible. This may result in residual confounding that should be considered when interpreting these results.”
Though we focused on a grouping of IMIDs where TNF-a is implicated, and TNF-inhibiting drugs are FDA approved for use, we did not measure TNF-a. Our results therefore do not allow us to prove or disprove an association between TNF-a and AD; we show only that in a group of IMIDs where TNF-a is implicated, there was no increased risk of AD.
Though in general the proportional hazards assumption is violated when Kaplan-Meir survival curves cross, in our case, the few number of respondents in the IMID group resulted in a blocky survival curve relative to the smooth curve in the significantly larger non-IMID group. Therefore, some visual crossing of survival curves is highly likely. We conducted extensive testing of the proportional hazards assumption and found no violations.
The benefit of using secondary data sources for a longitudinal analysis is the savings in time and cost and the lack of difficulties in study recruitment and retention common in clinic-based research. The tradeoff is a lack of validity in the diseases under study, as discussed above. We believe our research gives reason to be cautious in interpreting other studies showing an increased risk of AD in immune-mediated inflammatory diseases. We recommend future research pursue sources of validated diagnoses of IMIDs and AD to examine this relationship further. We also suggest that future research include sensitivity analyses for IMID classification when using administrative data.