Questions from Committee 1) Why did you perform 7:1 matching? We extracted all available cases of dementia meeting our criteria. A 7:1 matching was used to maximise power, particularly for rare exposures, while designed to keep within the number patients that CPRD allow for a single study. We have added text to the methods section as follows: “Seven controls were used to maximise power to detect associations with potentially rare exposures or covariates, while adhering to the limits of the data provider regarding maximum sample size for a single study.” 2) The effect size is marginal (upper bound of 95% CI is only 1.14) and the potential for residual confounding high (were these medications prescribed for disorders associated with dementia risk OR even for subtle symptoms of very early disease? Strength of effect size. The effect size referred to by the committee (OR=1.11 95% CI=1.08-1.14) reflects the use of ‘any’ ACB3 in the MEP vs none. We would not expect this particular effect to be large, because it averages the effect from those using less than 3 months ACB3 (where little or no effect is expected) with the smaller number using 3 months or more ACB3 (where a much larger effect is expected and indeed observed). The dose-response analysis also in table 2 is the more meaningful to judge the size of effect. We have added a section (see response to reviewer 2 comment 35 translating the effect size for anti-depressants into number needed to harm to aid clinical interpretation). Residual confounding We acknowledge that residual confounding can never be completely excluded in observational research, particularly in pharmacoepidemiology, and much of the strength of our report rests in the range of analyses presented specifically to test different ways in which confounding might have affected our results. We have added an extra appendix to further explore this possibility (see final comment in response to reviewer 3) and we believe that these demonstrate that residual confounding is unlikely to account for the association between ACB3 and dementia. In particular: • There is a strong dose-response relationship between exposure to ACB3 and dementia incidence (table 2). • There is little difference in the associations reported when we adjust for covariates measured before start or before the end of the MEP (table 3). • Several associations remain when we restrict only to those exposures occurring 15- 20 years before dementia diagnosis (table 4). • There is no independent effect of the main potential confounder (depression) in the multivariate analyses restricted to 15-20 years before dementia incidence or in the dataset overall (appendix 4). • If residual confounding by poorly recorded depression did account for our finding then we would expect to see less effect among those with severe depression recorded by their GP. Our new analysis shows that the effect of anticholinergic antidepressants does not depend on the presence or severity of recorded depression during that time (appendix 4). Conversely, these same analyses show that apparently significant association between ACB1 exposure and dementia incidence indeed be attributable to confounding. Here there is no dose- response effect, and the association between dementia and ACB1 medications does not persist when exposures more than 10 years before diagnosis are considered (table 4). We have added the following to the result section: “Appendix 4 details a further analysis of the associations between depression, antidepressants, and dementia incidence. Depression diagnosis has no independent association with dementia incidence after adjusting for antidepressants and other covariates in the model. However antidepressants were consistently associated with dementia, this effect was not attenuated by controlling for depression, and there was no interaction between depression severity and anti-depressant use.” 3) You allow for a 4 year period to avoid protopathic bias but the “latent” period for dementia may be longer than that.) We agree it is likely that the latent period for dementia before diagnosis exceeds four years in many cases. For this reason we conducted the analysis given in table 4 (stratifying exposure by time since index date), which essentially also tests lag times of 10 and 15 years. This shows that the effects of ACB3 antidepressants and urological medications persist even for exposures 15-20 years before dementia diagnosis. For protopathic bias to account for this result urinary incontinence would need to be substantially increased 15-20 years before dementia diagnosis. We have acknowledged this possibility in our discussion section as follows: “Lower urinary tract symptoms themselves have been linked to future dementia incidence (35) and may be a symptom of early neurodegeneration (36), but to account for our finding unrecorded urinary incontinence would need to be a substantial risk factor for dementia diagnosed between 15 and 20 years later.” 4) Some patients “without dementia” may be “incipient dementia or very early, as yet undiagnosed dementia and if we follow for long-enough they will have dementia”. This is another potential source of bias. We agree, as is the case with any study of chronic disease with insidious onset (eg hypertension, diabetes etc) many of our non-dementia control group may in fact have dementia that is not yet diagnosed, or have incipient cognitive impairment not yet severe enough for a diagnosis. Also, any misclassification would shrink apparent effect sizes toward null through regression dilution bias, although could not remove them completely or induce a negative correlation. An analysis restricting to controls surviving beyond the index date without dementia would introduce a positive bias associated with the requirement to survive beyond the index date, and in any case many patients with dementia are never diagnosed. We have added the following to the discussion section under ‘strengths and limitations’ “..Hence it is possible that some of our controls have undiagnosed dementia or early cognitive impairment. This misclassification would shrink estimated effects toward the null (OR=1) but would not remove them completely. “ 5) In the database, what is the severity of symptoms required to make a diagnosis of dementia? Primary care data only includes complete information on whether or not a patient has ever received a diagnosis of dementia, and the date on which this diagnosis was made. In the UK dementia diagnoses are most often made in secondary care by specialist memory services. We have added the following text to the discussion under strengths and limitations: “In the UK most dementia diagnoses are made by specialist memory services which are then communicated to primary care, and so our cases are very likely to represent genuine cases of dementia.” Further we have found in a separate (complete but unpublished) study comparing primary care diagnoses with objective measures of dementia status that almost all diagnosed cases in English primary care meet objective criteria for dementia when compared against diagnosis equivalent to DSM-III-R. Hence dementia patient in the UK would need to meet a threshold equivalent DSM-III-R to receive a diagnosis. 6) Many patients have mixed dementia, often with vascular etiology. Some patients may have had cognitive decline not yet coded as dementia and then suffered a stroke. We agree many patients will have mixed dementia. Clinical expression of dementia in an individual reflects contributions from different pathologies and so the effects of risk factors may well contribute to cognitive impairment and dementia irrespective of the ‘primary’ subtype. For this reason and the inconsistent way in which dementia subtypes are coded in primary care data we have not disaggregated our cases based on dementia subtype. 7) Not all editors were familiar with Dosulepin. Response: Dosulepin is a tricyclic antidepressant (used to be known as Dothiepin) we have added this information under ‘frequency of medication use’ in the results section. 8) The committee wasn’t sure of novelty. Novelty is addressed by the responses to comments 9,10 and 11. In short, previous studies have suggested that ‘anticholinergics’ are associated with dementia incidence. However this finding is much too broad to be of practical use and previous studies have been too small and without enough follow-up or power to control for confounders. The uncertainty around this finding has meant that previous findings have been easily dismissed and have not been acted on. We considerably narrow the groups of drugs around which concern should remain and provide much stronger evidence on their effect on dementia incidence. 9) What this seems to add is more granularity in terms of anti-cholinergic burden and drug class, and we find that the answer is not straightforward and there’s not an overall class effect. Do we need studies on subgroups of anticholinergics? First, ‘anticholinergics’ describes a very wide class of medications with different actions and effect on the brain. Previous studies have not disaggregated these effects by pharmacological action or indication. Second, it has not been possible to rule out confounding with prodromal symptoms or risk factors of dementia as rigorously as we have in this study. This is because previous studies have had shorter follow-up periods, smaller samples preventing disaggregation or poor confounding control owing to incomplete data on covariates. Finally, many drugs (used at any time by around 50% of the middle aged and older population) are currently considered to be ‘mildly’ anticholinergic, reflected here by our ACB1 category. As previous studies had summed these effects together with the better established potent anticholinergics it was not known whether and to what extent these mild anticholinergics posed a risk of dementia. A very common drug with a small increase in risk could have a potentially large impact on dementia occurrence. This along with the failure to disaggregate classes has led to the creation of anticholinergic burden measures that simply sum medications.