Effect of fluoridated water on invasive NHS dental treatments for adults: the LOTUS retrospective cohort study and economic evaluation

Systematic reviews

There have been eight systematic reviews of the effectiveness of water fluoridation in preventing dental caries since the year 2000. 43–51 The first was carried out in the UK by the York Centre for Research and Dissemination and was commissioned by the Department of Health. 43 This review was assessed as high quality using the A MeaSurement Tool to Assess Systematic Reviews (AMSTAR 2) checklist for systematic reviews. 52 For the caries-effectiveness research question, 23 controlled before–after studies, 2 prospective and 1 retrospective cohort studies were included. Most studies demonstrated a protective effect of water fluoridation on dental caries in children, but the quality of the studies was low, mainly due to lack of consideration of confounding or effect modifying factors. As a result, in 2002, the UK Medical Research Council (MRC) was tasked by the Department of Health to identify priorities for future research. The recommendations of the working group included:

• the effect of fluoridation on the dental health of adults

• the impact of water fluoridation against a background of widespread topical fluoride use

• economic impacts and the effects on health and well-being beyond the usual measures of decayed, missing and filledteeth (DMFT)

• the effect of fluoridation on social disparities in dental caries.

The most recent systematic review to carry out a meta-analysis was conducted by Cochrane Oral Health in 2015. 47 The meta-analysis of 10 studies with 39,966 participants estimated that the initiation of water fluoridation was associated with a 26% relative reduction in the number of permanent teeth affected by decay [DMFT was reduced by 1.16 teeth in the fluoridated groups (95% CI 0.72 to 1.61)]. 47 Only prospective studies, with a concurrent control, and outcomes evaluated at two time points (before and after a change in fluoridation status) were included. 47 The quality of this review was assessed as ‘high’ using the AMSTAR 2 critical appraisal tool for systematic reviews. 52 The review authors had limited confidence in the size of the summary effect estimate due to the high risk of bias in most of the included studies. 47 The main issues were lack of consideration of confounding factors and non-blind outcome assessment. Furthermore, most of the included studies were carried out prior to the widespread introduction of fluoride toothpastes; none of the studies included adult participants, and there was insufficient evidence to evaluate the effectiveness of water fluoridation in reducing social inequalities in dental caries experience.

Three of the eight systematic reviews have included some adult participants. 45,49,51 The first was by Griffin et al. in 2007, which aimed to investigate the effectiveness of a range of fluorides for preventing caries in adults (gels, rinses, toothpastes and water fluoridation). 45 For the assessment of water fluoridation, they identified eight cross-sectional studies and one prospective cohort study. 45 A meta-analysis of seven of the cross-sectional studies that had assessed the effect of lifetime exposure found water fluoridation was associated with a 34.6% relative reduction in the number of permanent tooth surfaces affected by decay [decayed, missing and filled surfaces (DMFS)] (95% CI 12.6% to 51.0%). The quality of this review was assessed as being ‘critically low’ using the AMSTAR 2 tool. 52 The main issues were that the review did not include any evaluation of the quality of the primary studies or whether confounding was adequately accounted for. Additionally, the fluoride concentration of three of the studies included in the meta-analysis was above the maximum permitted level of 1.5 mg F/l in the UK which limits the applicability of the findings to the UK context, and some of the included studies were conducted in the pre-fluoride toothpaste era when caries levels were much higher. Furthermore, relative percentages can be misleading, as they make the difference seem larger than it is in clinical terms, particularly with such a finely graded outcome as tooth surfaces.

Two further systematic reviews that included adult participants were carried out on behalf of the Australian National Health and Medical Research Council (NHMRC) in 201649,50 and the Canadian Agency for Drugs and Technologies in Health (CADTH) in 2019. 53 The NHMRC review was an update to an earlier 2007 review but with much broader inclusion criteria. 44 The inclusion criteria for both reviews permitted any study design evaluating water fluoridation at concentrations of between 0.4 and 1.5 mg F/l that had conducted multivariable analysis to control for confounding (confounders not specified). Both reviews assessed dental and other health outcomes and were part of a wider examination of water fluoridation which also included ethical considerations53,54 and, in Canada, cost effectiveness. 55 The Canadian review included all of the systematic reviews45,47,56 and primary studies identified by the Australian review, a 2016 systematic review of the cessation of water fluoridation48 and any relevant primary studies published after the search period of the Australian review. The quality assessment for the Australian review50 using the AMSTAR 2 tool is ‘moderate’ and ‘high’ for the Canadian review. 51 Both reviews presented their findings as narrative summaries rather than meta-analyses due to the heterogeneous nature of the included studies. 50,51

The primary studies which included adult participants identified by the NHMRC review assessed caries57,58 and tooth loss. 59–62 Additional adult studies included in the Canadian review assessed caries severity (DMFT/S),63–66 caries prevalence,67 tooth loss68,69 and dental service use. 70,71 Of these 15 primary studies, 9 were assessed as being of ‘acceptable’ quality by the review authors and 7 as ‘low’ quality. 49,51 The studies were ecological57–62,65,67,70,71 and cross-sectional. 64–66,68 The authors of the Australian review concluded that the quality of the evidence was ‘low or very low’, but it was consistent and ‘indicates that water fluoridation improves the dental health of children and adults’. 50 The Canadian review also concluded that while the quality of the evidence was mixed, there was ‘consistent evidence that water fluoridation reduced caries in children and adults’, despite ‘widespread availability and use of fluoride toothpastes’. 51 With regard to other dental outcomes that included adult participants, the Canadian review authors found insufficient evidence for an association between water fluoridation and dental service use, limited evidence for reducing tooth loss and limited evidence for a reduction in social disparities. 51

In their recommendations for future research, the authors of the Canadian review noted that while the majority of studies identified for the caries outcomes were ecological or cross-sectional, these designs may be sufficient as long as appropriate adjustment for confounding is undertaken. 51 This is a view echoed by others who have highlighted that the quasi-experimental study designs specified as an inclusion criterion for the Cochrane systematic review were only possible during the initial implementation phase of water fluoridation (i.e. controlled before–after evaluations). 57,72 Most water fluoridation programmes were implemented in the 1950s and 60s, limiting the opportunities for such studies in more recent times. Adequately controlled cohort or cross-sectional studies have been suggested as more appropriate study designs in contemporary populations where it may not be possible to evaluate a change in water fluoridation status. 57,72 The Canadian review authors recommended that future research assess the impact of water fluoridation on reducing disparities in oral health and include populations that are varied in age, socioeconomic status, ethnicity and migration status. 51 The importance of conducting studies that can be generalised to the relevant national context was also noted.

Recent primary studies of the effectiveness of water fluoridation in adults and adolescents

The studies reviewed in this section are those 15 studies with adult participants, which were included in the 2019 CADTH review, as well as 4 studies published after the search end date of that review (December 2018). The search terms used to identify recent primary studies not included in the CADTH review were (Adult or Adolescent or Older or Elderly) AND (Caries or Decay or Tooth Loss or Dental Treatment) AND water fluorid$. The full search history is available in Appendix 1. The databases searched included: MEDLINE, EMBASE and all evidence-based medicine reviews (includes Cochrane and NHS Economic Evaluation Database). The search time frame was 1st of Jan 2018 to 21st of June 2022.

The identified studies were published between 2009 and 2022. The majority were carried out in Brazil59,61,64,67–69,73,74 and Australia. 57,58,60,63,65,66 All of the studies were of cross-sectional or ecological design. Most of the Brazilian studies were secondary analyses from the 200359,61 and 201067,68,73 SB Brazil national surveys of oral health, while two were the same cohort of participants at different time points64,69 and one was a school-based sample. 74 Four of the Australian studies were secondary analyses of the Australian National Survey of Adult Oral Health (ANSAOH) 2004–6,57,60,63,66 while one examined young offenders58 and one was an opportunistic follow-up of a childhood study sample. 65 Three of the studies utilised dental service data from South Korea,71 the USA75 and Sweden. 70 One study included Canadian Army recruits,76 and one was a secondary analysis of data from the telephone-based US Behavioural Risk Factors Survey 1995–9. 62 The outcomes assessed included self-reported number of lost teeth,62,69 clinically examined number of lost teeth,59,61,68,73 clinically examined prevalence and severity of decay (DMFT/S)57,60,63–67,74,76 and dental visits and/or treatments received. 70,71,75

All of the Brazilian studies reported that greater exposure to water fluoridation resulted in improved oral health, and the size of the differences between fluoridated and sub/non-optimally fluoridated groups was relatively large in comparison to the studies from Australia or the USA. For example, in the SB Brazil 2010, adults aged 35–44 living in sub/non-optimally fluoridated areas had a 27% higher mean ratio of missing teeth (MR 1.27, 95% CI 1.11 to 1.45). 73 In absolute terms, this was a large difference of three teeth per person, comparing the fluoridated and sub/non-optimally fluoridated municipalities,73 reflecting the high prevalence of missing teeth in this population. Another study which examined the SB Brazil 2010 data for 12–19-year-olds found that those living in sub/non-optimally fluoridated municipalities had 42% increased odds of having at least one decayed tooth (OR 1.42, 95% CI 1.08 to 1.86) and 57% increased odds of having at least one missing tooth (OR 1.57, 95% CI 1.16 to 2.14). 67 Both of these studies benefitted from calibrated examiners and a multilevel analysis which included a range of individual-level and contextual variables. 67,73

The Australian studies all found that increased exposure to water fluoridation resulted in fewer teeth affected by decay (DMFT/S) in younger adult age groups, but there were conflicting results for older adults. 57,58,60,63,65,66 The studies by Slade et al. (2013) and Do et al. (2017) both used data from the complete age range of participants in the ANSAOH 2004–6, but took a different analytical approach and reported different results. Slade et al. found those born pre-1960 who had more than 75% lifetime access to fluoridated water had 10% lower DMFT and 30% lower DFS (note ‘M’ component excluded) compared to those with < 25% lifetime access. In the cohort born 1960–90, for the same comparison, DMFT was 11% lower and DFS was 21% lower. 57 The estimates for the pre-1960 cohort are at risk of residual confounding due to age because of the relationship between per cent lifetime access to fluoridated water and age. That is, those in the pre-1960 cohort with > 75% lifetime access to fluoridated water were younger than those with a lower per cent lifetime access (due to the timing of water fluoridation implementation across Australia), but adjustment for age was only made by decade. Therefore, the analysis by Do et al. is preferred. 66

In the Do et al. study, adults aged 15–34 years of age who had the highest quartile of per cent lifetime access to fluoridated water had a 33% lower mean ratio of DMFS, compared to those with the lowest quartile of lifetime access (MR 0.67, 95% CI 0.48 to 0.92). 66 In the 35–44 age group, the mean ratio of DMFS scores was 22% lower (MR 0.78, 95% CI 0.66 to 0.93), but no difference was found in the age groups 45–54 and over 55. 66 In absolute terms, this is a difference of 3.1 tooth surfaces per person in the 15–34-year-olds and 5.7 surfaces per person in the 35–44-year-olds. These estimates were adjusted for age, sex, residential location, dental visit pattern, toothbrushing frequency, household income and oral hygiene. 66 The authors proposed that the absence of effect in the older age groups may be due to lack of variation in water fluoridation exposure in this age group and saturation of the DMFS score due to new caries being most likely to occur on previously treated surfaces (which would not increase the DMFS count). The main quality issue affecting the studies using ANSAOH data is that a large proportion of the original participants were excluded as a result of missing residential history information (26–31%, depending on the study). 57,66 An Australian follow-up study of young adults who had taken part in an earlier childhood study found that 20–35-year-olds with 0–74% lifetime access to fluoridated water had 26% higher rate of DMFS compared to those with 100% lifetime access [incidence rate ratio (IRR) 1.26, 95% CI 1.01 to 1.57],65 a difference of 1.6 tooth surfaces per person in absolute terms. 65 Finally, an Australian study of young offenders aged 13–21 found that those whose usual place of residence was sub/non-optimally fluoridated had a 77% higher rate of DMFT compared to those who usually resided in fluoridated areas (IRR 1.77, 95% CI 1.11 to 2.83). In both these studies, estimates were adjusted for important confounders such as age, sex and measures of socioeconomic status. 58,65

The Canadian study of Army recruits aged 16–45 found those recruited from areas with water fluoridation had a mean rate of DMFT that was 11% lower than the mean in recruits from sub/non-optimally fluoridated areas, after adjusting for age and gender [a difference in means of 0.67 fewer teeth (95% CI 0.79 to 0.55)]. 76 The mean rate of DMFS was 14% lower in those recruited from fluoridated areas [an absolute difference in means of 1.77 fewer tooth surfaces per person (95% CI 2.09 to 1.46)]. This study also evaluated if the effect of water fluoridation differed according to county-level income and education. No significant differences were identified, but the study population may not have had enough variation in terms of county-level income and education to fully examine this as all participants were Army recruits. 76 The US study of self-reported tooth loss in adults found those with a 100% probability of being exposed to water fluoridation at birth had 0.26 more teeth per person than those with a 0% probability (based on year of birth, county-level implementation date and population coverage of water fluoridation). 62 When stratified by social factors, the difference in the number of teeth was larger for black people (0.37 more teeth per person), those without a high-school diploma (0.61 more teeth per person) and high-school graduates (0.39 more teeth per person), whereas it was smaller for white people (0.18 more teeth per person) and was no longer statistically significant for college graduates (0.06 more teeth per person). 62 A range of county and individual-level covariates were included in the models for these estimates, including age, gender and education. 62 This is the only identified adult study to suggest water fluoridation may have a larger effect in more disadvantaged groups. 62

Three studies used routinely collected dental service monitoring data. 70,71,76 A South Korean study used National Health Insurance data to explore the effect of regional water fluoridation status in 2013 on dental outpatient visits between 2003 and 2013. 71 Patient records were obtained from a randomly selected National Health Insurance Sample Cohort. Patients were included in the analysis if they lived in non-metropolitan areas and had not had a caries diagnosis prior to 2003. The analytical sample was 472,250, and the average years of follow-up time was 9.12 years. Confounders included patient age, sex, income, health insurance funding bracket and dentist: population ratio and local government financial independence rate. The adjusted models found that patients living in fluoridated regions had a 5% lower risk of dental care visits (HR 0.95, 95% CI 0.93 to 0.97) over 9.2 years. The mean survival time to the first diagnosis of dental caries was 16 days longer in fluoridated regions (p < 0.0001). In those patients who did have a dental care visit, the number of visits per person was marginally lower in the fluoridated regions (β = −0.029, p = 0.04), but there was no difference in the costs of dental care. 71 This study illustrates that with the very large sample sizes afforded by routinely collected data sets, even very small differences will be statistically significant.

A Swedish study of adults aged 16 and older investigated the effect of lifetime water fluoridation exposure on aggregated area-level dental service data for 2008 and 2013. 70 The authors estimated in a regression analysis that if the fluoride level in a municipality increased by 1 mg F/l, the proportion of 20-year-olds who had a tooth repair in any given year would reduce by 0.6 percentage points (p < 0.01). 70 The regression included data from Swedish registers to allow for adjustment of confounders including sex, marital status, parent’s education, parent’s income, father’s cognitive and non-cognitive ability, parental immigration status and cohort mean education. A study of Medicare claims for children and adolescents aged 0–18 in Alaska was conducted to provide a comparator75 for an earlier study where water fluoridation had ceased. 77 The earlier study found that after the cessation of water fluoridation, the mean number of caries-related dental procedures per child aged 13–18 increased by 0.29 from 2003 (pre cessation) to 2012 (post cessation), and the costs increased by $145 per child. 77 The comparator study found that in a neighbouring area with similar census characteristics and ongoing water fluoridation, the mean number of caries-related dental procedures per child aged 13–18 decreased by 0.12 compared to 2003 and 2012, and the costs per child increased by $68.61. 75 The pre–post estimates in both studies were adjusted for sex, race and inflation, but as the studies were reported separately, no consideration of the effect of differences between areas was possible.

In summary, recent studies of the effectiveness of water fluoridation in adolescents and adults have found that water fluoridation is associated with improved oral health in adults aged up to 45 years. Most of the studies identified conducted multivariable analysis with consideration of confounders including a minimum of age, sex and socioeconomic status; only two studies did not. 75,76 However, much of the recent evidence comes from contexts which are not comparable to the UK and may have higher levels of dental disease and unmet treatment needs, including those studies conducted in Brazil59,61,64,67–69,73,74 and those with specific population subgroups such as young offenders or Army recruits. 58,76 The studies which are most applicable to the UK general population (i.e. Australia, the USA and Sweden) have demonstrated absolute differences so small that whether they are meaningful from the perspective of patients, public and clinicians is debatable. Additionally, there remains little evidence of the effect of water fluoridation on the oral health of older adults and its role in reducing social inequalities in adult oral health.

The NIHR CATFISH study: an evaluation of a water fluoridation scheme in Cumbria

A landmark 7-year prospective cohort study on the effectiveness and cost effectiveness of water fluoridation for preventing dental caries in children has recently been completed in the UK and is mentioned here because it is a contemporary, large UK study funded by National Institute for Health and Care Research (NIHR) and therefore highly relevant to the LOTUS study. 78 This study was made possible because of the replacement of water fluoridation dosing equipment after several years of inactivity. This created an opportunity to study the effects of reinstating the programme as if it were a new scheme. The study design was a controlled before-and-after study, with the start of the study being < 1 year since the change in fluoridation status. Two cohorts of children were followed, one from birth to 5 years and one from 5 to 11 years. The two cohorts allowed the effects of fluoride at different time points to be tested. There has long been debate over whether there is an important systemic effect of water fluoridation or whether the effect is primarily topical. For a systemic effect, fluoride must be present in the tissues when the teeth are developing in the jaws. For the deciduous (baby) teeth, this occurs from 30 weeks in utero up to 12 months of age. 79

This study found that, after adjusting for deprivation, age and sex, 5-year-old children who were exposed to fluoridated water from before birth (n = 1444) had 26% reduced odds of developing decay [odds ratio (OR) 0.74, 95% CI 0.56 to 0.98]. In absolute terms, the prevalence of any decay was 17.4% in children who had received fluoridated water and 21.4% in children who received sub/non-optimally fluoridated water. For the older cohort (n = 1192), while there was a lower prevalence of decay in the fluoridated group (19.1%) compared to the sub/non-optimally fluoridated group (21.9%), the 95% CI for the OR included the possibility of no effect (OR 0.80, 95% CI 0.58 to 1.09). There was no evidence of water fluoridation reducing the effect of social deprivation on either the prevalence of decay or the number of teeth affected by decay (DMFT, secondary outcome). The economic evaluation found that in the birth cohort there was a high probability of water fluoridation being cost-effective at a willingness-to-pay threshold of £20,000 per quality-adjusted life-year (QALY).

For comparison of effect sizes, the 2015 Cochrane systematic review meta-analysis estimated an absolute reduction in the prevalence of decay in deciduous teeth of 15% (95 CI 11% to 19%). 47 The Cumbrian Assessment of Teeth a Fluoride Intervention Study for Health (CATFISH) study authors concluded both the prevalence of decay and the size of the effect of water fluoridation were much smaller in this contemporary UK population than previously reported, and the clinical relevance of such a difference would need to be considered by policy-makers and the public. This study was designed and conducted in line with guidance on quality in water fluoridation studies and is highly relevant as it was carried out in the UK in a contemporary population. However, the shifting epidemiology of tooth decay into older age groups and the chronic long-term progression of this disease throughout life may mean the cohorts of young children included in this study were yet to experience the greatest risk of disease and therefore the greatest potential benefit of water fluoridation.

Participants

Adults and adolescents aged 12 years and over, attending NHS dental practices in England in the last 10 years (prior to NHS BSA selecting the study cohort from within the database).

Adolescents aged 12 years and over were included as well as adults, as this is the age at which the permanent, adult teeth are usually present in the mouth (except for third molars or ‘wisdom teeth’). This age was also chosen to provide a useful continuation of information, following on from the cohort included in the (at the time, ongoing) NIHR PHR CATFISH study of water fluoridation in children.

Inclusion

1. Dental records that can be assigned to a unique individual using the combination of NHS BSA identifier (initial, surname, gender, date of birth) and NHS number.

The inclusion criteria were important because of the way that these data are stored within the NHS BSA database (as discreet claims relating to single courses of treatment rather than patients). The older claims (2010–5) in the NHS BSA database did not include the patient’s NHS number but rather used an NHS BSA patient identifier (made up of a combination of surname, first initial, date of birth and sex). These identifiers cannot be considered as unique at the national level, therefore a one-to-one match between BSA ID and NHS number was required to ensure all claims related to a single individual. Around 70% of the claims submitted into the new (post 2015) NHS BSA database can be assigned an NHS number. Further detail on this linkage process is provided in Data linkage and validation.

Exclusion

1. Individuals were excluded from further analysis if they did not have at least two episodes of dental attendance within the 10-year observation period.

The reason for this exclusion criteria was that we wanted to have at least two time points for information on residential address to strengthen the assignment of exposure to water fluoridation.

Source population

Users of NHS primary care dentistry services in England and Wales. This includes data on patients attending a range of different types of dental services: general dental services (GDSs; high-street practices), community dental services, prison dental services, urgent and out-of-hours services, specialist primary care referral services (oral surgery/sedation/orthodontics) and domiciliary care (for housebound patients).

Database population

Users of NHS primary care dentistry services in England and Wales between 22 November 2010 and 21 October 2020.

The retention period for NHS BSA dental data is 10 years, and the data were downloaded on 21 October 2020. The study team did not have access to the NHS BSA database population. The study population was selected by NHS BSA.

Study population

Users of NHS primary care dentistry services in England and Wales between 22 November 2010 and 21 October 2020, meeting the study inclusion and exclusion criteria, as well as NHS BSA selection criteria (Figure 2) and transferred to The University of Manchester. As per the s251 support from the Health Research Authority’s Confidentiality Advisory Group (CAG), all patients with an NHS data opt-out were excluded from the study population by NHSSA. NHS BSA also removed any treatment claims that related solely to orthodontic care.

FIGURE 2.

Summary of the pathway into the study population.

A schematic illustration giving an overview of pathway for inclusion in the study population is provided in Figure 2.

Water fluoride concentration data

Routinely collected drinking water quality monitoring data were used to create the patient-level fluoride exposure variable.

In the original research proposal, we planned to use a binary postcode look-up table detailing which areas in England are ‘flagged’ as having a water fluoridation programme. We requested data on the coverage of water fluoridation and water fluoride concentrations from the Drinking Water Inspectorate (DWI) and Public Health England (PHE). Neither PHE nor DWI was able to provide a binary indicator of the coverage of water fluoridation programmes at a small area level, and they advised us to contact the water suppliers directly. We contacted the 24 water suppliers listed by the water regulator (Ofwat) in England and Wales. 83 The water companies did not have postcode-level mapping of fluoridation status or an indicator of which areas were intended to be included in a water fluoridation programme. Therefore, it became necessary to map water fluoride concentrations ourselves using water quality samples for fluoride.

Water quality regulations in England and Wales stipulate that water companies must monitor the concentration of fluoride to ensure it does not exceed the maximum permitted value of 1.5 mg F/l. 84 The drinking water network is mapped using water supply zones (WSZs) rather than standard UK census-based geographies. A WSZ is an area within which the water supply is of a ‘similar nature and treatment’ and the population supplied does not exceed 100,000. 84 Water quality samples are drawn from randomly selected households (customer taps/faucets) within a WSZ, or, if the nature and treatment of the water are the same, they can be taken from supply points higher up in the water distribution network. The number of samples that are taken in each WSZ in each year is correlated with population size, with more frequent samples taken in more populous areas. 84,85 The standard sampling frequency for fluoride is 8 samples per year, per 5000 population, per WSZ. 84 Minimum standards apply in relation to the sampling methodology and the accuracy of water quality samples (ISO 17025 and the Drinking Water Testing Specification). 86

We requested sample data on water fluoride concentration (mg F/l) for every WSZ, for every year from 2009 to 2020 under the provisions of the Environmental Information Regulations (2004). Requests were made through the ‘What Do They Know’ website87 and via direct e-mails between March 2020 and November 2020. We requested that the sample data be supplied as Microsoft Excel files containing supply zone names and zone codes. When consumer tap samples were unavailable, we requested supply point samples of water fluoride data and a corresponding list of WSZs supplied by these supply points. We sought raw sample data where possible, rather than the annual summaries. To enable the WSZ data to be mapped onto standard UK census-based small area geographies [lower super output areas (LSOAs)], we also requested geographic information system (GIS) shapefiles (maps) of the WSZ boundaries for every year from 2009 to 2020. GIS shapefiles show the WSZ boundaries as polygons. A polygon is a series of X (easting) and Y (northing) co-ordinate pairs that enclose a geographical area.

In total, we submitted 18 requests through the ‘What Do They Know’ website87 and sent 74 direct e-mails. In some cases, our initial requests were refused, with reasons cited including concerns about the potential for terrorist attacks on the water supply, samples being considered as the personal data of consumers and unreasonable burden or information not held. In such cases, we then proceeded to request an internal review and, in one case, made a referral to the Information Commissioner’s Office. One water company insisted that we sign a non-disclosure agreement to ensure that we did not publish the data in relation to WSZ geographies, whereas the others supplied the data as publicly available, in line with the EIR provisions.

Dental treatment and oral health data

Routinely collected NHS BSA dental treatment claims data were used to provide the study population and outcomes data.

These data were collected by the NHS BSA from primary care dental practices contracted to provide NHS dentistry within England and Wales between October 2010 and November 2020. NHS BSA is an executive non-departmental public body of the Department of Health and Social Care (sometimes known as an arms-length body). The BSA is responsible for monitoring dental patient charges, processing clinical activity against dental contracts and making payments to dentists in line with their NHS dental contract arrangements. The BSA also has a role in monitoring population-level access to dental services and guarding against fraudulent claims. 88

Data on the number and type of NHS dental treatments provided during each course of treatment are sent to NHS BSA within 2 months of the completion of a course of treatment. A ‘course of treatment’ is defined as all treatment that is deemed to be clinically necessary and which the patient consented to after each new dental examination. The claims for each course of treatment are submitted to the BSA using a form known as the ‘FP17’. These claim forms are completed electronically, and the required data are automatically populated using information from the dental electronic patient record (EPR) software.

National Health Service dental care

National Health Service dental care includes any treatment the dentist believes is clinically necessary to achieve and maintain good oral health. 89,90 It does not include treatments that are purely cosmetic, for example, teeth whitening. It also does not include more aesthetic treatment options where a conventional treatment option would be better from the perspective of maintaining oral health. An example of this would be a bonded porcelain crown on a molar tooth that is more aesthetic but involves removing more tooth tissue than a conventional (thinner) metal crown. Patients may opt to pay for cosmetic treatments or more aesthetic treatment options on a private basis, and it is permitted for dentists to mix NHS and private treatments on the same patient as part of the same course of treatment. Private dental care is not reported to NHS BSA and is therefore not recorded in the NHS dental treatment claims database.

Types of dental services

There are a range of different types of primary dental care services which submit FP17 treatment claims to the NHS BSA. Patients attending the following types of services are included in the study population data set.

National Health Service contracting and payment systems for primary care dentistry

National Health Service dental care is provided by independent contractors who hold a contract to provide an agreed scope and volume of NHS dental care. The three types of contracts that are currently in use are described below.

1. General dental services contract 2006. 91

   This is the most frequently used type of dental contract and is the type used by most ‘high street’ general dental practices. 93 Contract holders must provide all ‘mandatory services’ and this clinical activity is commissioned and monitored using units of dental activity (UDAs). 91 The contract holder receives an agreed annual financial payment (total contract value) in return for providing a pre-specified number of UDAs. The total contract values and the target number of UDAs to be delivered are negotiated individually for each practice and are based on historical funding and activity levels. 94 Therefore, the financial value of 1 UDA varies between contracts (practices). The mean value of 1 UDA for GDS contracts in the financial year 2020–1 was £28.64. 95

   When this contract was introduced in 2006, over 400 individual treatment fee-for-item treatment codes were replaced with just three ‘bands’ of treatment complexity for each course of treatment:

   • A band 1 course of treatment (examination, prevention, radiographs) attracts 1 UDA.

   • A band 2 course of treatment (all care in a band 1 course of treatment plus any number of direct restorations, endodontics and/or extractions) attracts 3 UDAs.

   • A band 3 course of treatment (all care in a band 1 and/or band 2 course of treatment plus any laboratory work including indirect restorations and dentures) attracts 12 UDAs.

   The UDA system does not consider the number of treatments provided within each band. 94 For example, a course of treatment containing one direct restoration accrues the same number of UDAs as a course of treatment containing three direct restorations, two molar endodontic treatments and an extraction.

   It is important to note that practices working under this contract may also provide services that are commissioned and monitored using metrics other than UDAs. For example, orthodontics, domiciliary care, sedation, dental public health or advanced mandatory services. Orthodontic courses of treatment are commissioned using units of orthodontic activity (UOAs), and for other services, specific key performance indicators (KPIs) may be outlined in a service level agreement accompanying the main contract. Example KPIs might include the number of clinical sessions to be provided, the number of patients to be treated or the maximum waiting times. This feature means it is not possible to determine the real-term value of contracts using only the total contract value and number of commissioned UDAs.

2. Personal dental services (PDS) contract 2006. 96

   This type of contract is more commonly used for more specialised dental services that provide care on referral and even some secondary care services, such as treatment under general anaesthesia. 97 Services working under this type of contract may have a nominal number of UDAs to deliver, but they are more likely to be commissioned and monitored according to KPIs.

3. Personal dental services plus (PDS+) contract 2009. 96

This type of PDS contract was introduced with a view to increasing access to dentistry in underserved areas. Its payment mechanism is comprised of three payments in relation to services, access and performance. There are only 36 PDS+ contracts in England. 93

National Health Service dental patient charges

Patients in England pay for NHS dental care unless they are exempt from fees for a specified reason. In 2022, the patient charges are as follows:90

• Band 1 (examination, radiographs, prevention, treatment plan): £23.80

• Band 2 (band 1 treatments plus periodontal treatment, direct restorations, endodontics, extractions): £65.20

• Band 3 (band 1 and/or band 2 treatments plus indirect restorations and dentures): £282.80

• Urgent (temporary treatment to resolve acute pain or infection): £23.80.

The charge paid by the patient is based on the highest-complexity band that was provided during each particular course of treatment. It also covers any care provided in the less complex bandings. For example, a patient who had an examination, radiographs, two extractions and a partial denture would pay the band 3 charge of £282.80.

Patients who are currently fully exempt from NHS dental payment charges include those aged under 18, those aged under 19 and in qualifying full-time education, those who are pregnant or have given birth within the last 12 months, prisoners and those receiving qualifying low-income-related benefits or holding an income-assessed exemption certificate (HC2). 98 Some patients may be entitled to reduced NHS dental charges if they hold a partial NHS exemption certificate (HC3) through the NHS low-income scheme.

Data linkage and validation to create the study population

The following is based on information provided by NHS BSA and describes their standard processes for data linkage and validation, followed by the additional steps that were necessary to provide the study population data set.

National Health Service number (a unique patient identifier) is infrequently populated on FP17 claims forms when they are submitted by dental practices, as it is not a mandatory component. In October 2019, only 6.8% of claims processed by NHS BSA had an NHS number recorded on submission. As a result, NHS BSA undertake a process of identification in order to link dental treatment claims for separate courses of treatment to an individual patient.

Since the creation of a new NHS BSA dental data warehouse in February 2016, the standard process to link FP17 claims to a unique individual patient is for all newly submitted claims to be checked using the following identifiers:

• Surname

• Forename

• Sex

• Date of birth

• Last known postcode (patient)

• Dental contract ID.

If this combination of details has not been recorded previously, the patient’s demographic details are sent to the Demographic Batching Services Spine (DBS) to trace the patient’s NHS number. If the FP17 claim form was submitted with an NHS number and the traced NHS number differs, the traced NHS number is used. Around 70% of the patient records in the new data warehouse have an associated NHS number.

To provide data that spanned 2010–20 for this study, NHS BSA needed to link patient records in the current data warehouse with an older, legacy data warehouse. Demographic tracing of NHS numbers did not take place for records held in the pre-2016 data warehouse; therefore, NHS numbers are not usually available in the older system.

Rather, patients were identified using an Index ID made up of a combination of surname, first initial, date of birth and sex. Index IDs cannot be considered as unique identifiers at the national level. To link the FP17 data across both databases, the Index IDs were matched to the post-2016 patient records. Only patients who had a 1 : 1 match between their NHS number and their Index ID were selected for the study population to ensure that the treatment claims were associated with a unique individual.

FP17 records for each selected patient were then validated. FP17s were removed where there was no information entered, where multiple forms had been submitted for one claim and where the claim was later withdrawn. Courses of treatment that contained only orthodontic treatment were removed.

Non-differential misclassification

Another potential source of bias in this study is non-differential misclassification. This type of bias is caused by imprecision and errors in the measurement and classification of exposure and outcome variables. This could occur if a participant in our study was misclassified as being in the optimally fluoridated or sub/non-optimally fluoridated group, or if the dentist did not record the DMFT or numbers of treatments correctly. Non-differential means this bias is expected to affect both the optimally fluoridated and sub/non-optimally fluoridated groups equally. In situations where non-differential errors of misclassification of a binary exposure and outcome are independent (i.e. not connected, obtained from different sources, as in the present case), the effect of this type of bias is to obscure the treatment effect through random noise or ‘bias towards the null’. 111

Exposure: patient fluoridation status

Individuals within the NHS BSA data set were assigned a unique water fluoride concentration (mg F/l), based on the number of years they were assumed to have lived within each LSOA and the recorded water fluoride concentration for each LSOA. For further detail on how the 10-year grand mean water fluoride concentration was calculated for each patient, see Estimation of patient drinking water fluoride concentration.

The 10-year grand mean water fluoride concentration for each patient (mg F/l) was recategorised from a continuous into a binary exposure variable:

• Optimally fluoridated: 10-year individual grand mean water fluoride concentration of ≥ 0.7 mg F/l.

• Sub/non-optimally fluoridated: 10-year individual grand mean water fluoride concentration of < 0.7 mg F/l.

The threshold of 0.7 mg F/l and above was used to define ‘optimally fluoridated’ as a pragmatic cut-point because previous research has demonstrated that the 1 mg F/l target for water fluoridation programmes in England is not consistently achieved over the long term. 103,112 Furthermore, this is in line with the approach taken in the statutory 2018 health monitoring report for England, which used 0.7 mg F/l as the lower boundary in their binary analyses of caries-preventive effects. 112

Creating a binary exposure variable using continuous data will always present the difficulty that some individuals will be very close to the cut-point. Previous work by PHE found that between 2005 and 2015, 86% of the population of England lived in areas with low water fluoride concentrations (< 0.1 < 0.4 mg F/l), 4% lived in areas with between 0.4 < 0.7 mg F/l fluoride and 10% lived in areas with a water fluoride concentration ≥ 0.7 mg F/l. 112

Variables used in the matching process

Guidance on propensity score matching states that all confounders and important predictors of the outcome that are unrelated to exposure should be included in the propensity score model. The latter predictors were included to increase the precision of the effect estimates rather than to reduce bias. 110 The following confounding and predictor variables were identified through the creation of a DAG. 113–115

Confounders

The variables below are ‘surrogate’ confounders, as it is not possible to match on the ‘true’ confounders of historic regional prosperity and past population oral health. Surrogate confounders are variables which when conditioned on (through matching, stratification or adjustment) block the confounding pathways caused by the true confounders. 105

• Patient age

The age of the patient at the start of each course of treatment was added to the study by NHS BSA using the patient’s date of birth (prior to de-identification of the data). For our analysis, we assigned all patients an estimated ‘age in 2011’ as this was the first full year of data collection. This was created by adding or subtracting the number of years before/after 2011 from the age of the patient as recorded at their first course of dental treatment. Age was used as a continuous variable in the analysis.

• Patient sex

Patient sex, as recorded at their first course of dental treatment, was used for the analysis. This is recorded on the FP17 as a binary male/female variable.

• Patient ethnicity

The patient ethnicity variable, as supplied by NHS BSA, contained 18 categories at every course of dental treatment. These 18 categories were grouped into 6 categories as per 1–6 below. ‘Patient declined’ or ‘N/A’ were the most common responses at any single course of treatment. In cases where the patient’s ethnicity was recorded as ‘patient declined’ or ‘N/A’, ethnicity information from the next subsequent visit that was not ‘patient declined’ or ‘N/A’ was used to populate this variable.

1. White: White British, White Irish, other white background.

2. Mixed/multiple ethnic group: white and Black Caribbean, white and Black African, white and Asian, other mixed background.

3. Asian/Asian British: Asian or Asian British Indian, Asian or Asian British Pakistani, Asian or Asian British Bangladeshi, other Asian background, Chinese.

4. Black/African/Caribbean/Black British: black or Black British Caribbean, black or Black British African, other black background.

5. Any other ethnic group: any other ethnic group.

6. No information: patient declined or ‘N/A’ at all courses of treatment.

• Patient deprivation [Index of Multiple Deprivation (IMD) 2015]

Area-level deprivation in England is assessed using a standardised methodology which ranks the 32,844 neighbourhood areas (LSOAs) in terms of relative deprivation, with 1 being the most deprived and 32,944 being the least deprived. Deprivation is assessed across seven domains: income, employment, education, health, crime, housing and living environment. NHS BSA added the IMD 2015 deprivation rank to each course of dental treatment, using the patient’s home postcode. The 2015 IMD data were used rather than the 2010 IMD because the boundaries of LSOAs were reorganised in 2011. The data used to create the IMD 2015 relate to the tax year 2012–3. 116

We transformed the deprivation ranks to deciles, using the rank cut-points provided by the Department for Communities and Local Government, with decile 1 being the most deprived 10% and decile 10 being the least deprived 10% of neighbourhoods (LSOAs) in England. 117 For the analysis, we linked the 2015 deprivation decile to the patient using their LSOA of residence as recorded at their first course of dental treatment.

• Rural–urban classification (2011)

The Office for National Statistics (ONS) rural–urban classification is an official statistic which classifies LSOAs into eight categories, ranging from the most urban, ‘Major Conurbation’ to the most rural, ‘Village and Dispersed in a Sparse Setting’. 118 The patient’s LSOA of residence at their first course of treatment was used to assign a rural–urban classification to them.

• Dental practice deprivation (IMD 2015)

The LSOA of the dental practice the patient attended for their first course of dental treatment was used to assign the dental practice deprivation variable. As with patient deprivation, this used the 2015 IMD deciles supplied by the Department for Communities and Local Government.

• Local NHS dentistry provision (UDAs commissioned per person, by local authority)

The number of UDAs commissioned in the financial year 2015–6 by local authorities was supplied by NHS BSA in response to a request under the Freedom of Information Act (i.e. it was not derived from the main NHS BSA study data set). The year 2015–6 was the earliest year this information could be supplied for. For the analysis, we created the variable ‘UDAs commissioned per person, by local authority’ using the ONS mid-year 2015 population (all ages) as the denominator. The patient’s home LSOA at their first course of treatment was used to link ‘UDAs commissioned per person by local authority’, to individual patients using ONS LSOA-to-local authority look-up files.

Outcome predictors

The addition of the variables below was not deemed essential for the control of confounding. That is, these variables are in addition to the ‘minimally sufficient adjustment set’ and were added as outcome predictors (effect modifiers). 110

• Type of dental contract

Primary care dental practices in England and Wales are contracted to provide NHS care using three different types of contracts, which are used for slightly different types of primary care dental services and have different blends of monitoring metrics (see Selection of the study population by National Health Service Business Service Authority for further detail). Reflecting these different contract types, this variable had three categories:

1. General dental service

2. Personal dental service

3. Personal dental service plus.

The type of dental contract that the dental practice was working at the patient’s first course of dental treatment was used to assign this variable.

• Patient NHS Exemption Status

The reason for any NHS dental patient charge exemption at each course of treatment was supplied as a text variable by the NHS BSA. To create the patient’s NHS exemption status variable, the reasons for exemption were grouped into three categories, as listed below, and the patient’s exemption status at their first course of dental treatment was used.

1. Not exempt

2. Income-related exemptions: tax credit, income support, pension credit guarantee credit, NHS Tax credit exemption, Job seekers allowance, universal credit, income-related employment support allowance, full remission, HC2 certificate, partial remission, universal credit, income-based job seekers allowance, partial remission HC3 certificate, disability working allowance

3. Non-income-related exemptions: expectant mother, nursing mother, mother of child born in the last 12 months, 18 + further education, aged 18 in full-time education, patient under 18, prisoner, exam only, person of state retirement age.

Overview of propensity score matching

A key consideration in the design of this study was the selection of appropriate individuals residing in areas that were not exposed to water fluoridation to provide a comparator group in order to minimise bias and strengthen causal inference. In earlier observational studies, the selection of non-exposed individuals was sometimes carried out opportunistically, selecting individuals from neighbouring local authority areas or other, broadly comparable areas. We used propensity scores to formally select individuals in the sub/non-optimally fluoridated group based on their similarity to individuals in the optimally fluoridated group using a range of characteristics relating to exposure and outcome. A propensity score is a numerical value from 0 to 1 that reflects the probability of assignment to a particular study group (in this case, receipt or no receipt of optimal water fluoridation) based on their characteristics at the study baseline. 129 The characteristics the estimated propensity score is based on are those chosen by the study team as being important to consider to reduce confounding bias. In assigning each individual a propensity score, the overall ‘balance’ of the distribution of important characteristics in the exposed and non-exposed groups can be assessed using a single numerical value.

The aim is to achieve, as far as possible, an equal balance between the two groups in the pre-specified important characteristics upon which the propensity scores are based. 129 Once the propensity scores have been generated, individuals in the exposed (optimally fluoridated) group can be matched to those in the non-exposed (sub/non-optimally fluoridated) group based on the value of their propensity score. In this study, we used distance matching methods, whereby the distance (difference in propensity score for an individual in the exposed group and an individual in the non-exposed group) is used to form matched clusters of exposed and non-exposed individuals.

Guidance on covariate selection to estimate propensity scores106,129 suggests the best balance (and hence, reduction in confounding) is achieved by including baseline characteristics that, based on expert content knowledge:

• are related to both the treatment and the outcome (e.g. receipt of water fluoridation and receipt of invasive NHS dental treatments)

or

• are expected to be related to the outcome (receipt of invasive NHS dental treatments, DMFT and missing teeth)

and

• are not on the causal pathway between exposure and outcome.

In the case of a very large number of potential variables, it is recommended to include those baseline characteristics related to both exposure and outcome and discard those expected to be only weakly associated with the outcome. 106–109 To ensure potential covariates meet the above criteria, we created DAGs as described in Bias and Appendix 2, Tables 33 and 34 and Appendix 2, Figures 16–21. These DAGs illustrate the hypothesised relationships between all common causes of exposure and outcomes. 110

To inform the selection of the propensity score variables, we had planned to hold a 1-day workshop where key stakeholders would be invited to consider potential confounders that needed to be addressed during the analysis. However, due to COVID restrictions, this was changed to a series of online consultations. See Variables for details of which variables were chosen for matching. Further information on the selection of variables is included in Appendix 2 (Appendix 2, Tables 33 and 34 and Appendix 2, Figures 16–21).

The following sections describe the steps taken to create the propensity scores, match exposed individuals to non-exposed individuals forming a data set for analysis and estimating the effect of exposure on the clinical outcomes.

Data were validated and cleaned prior to any analysis. The extent of missing data for the exposure variable, covariates used to estimate the propensity score and study outcome variables were documented. No imputation was undertaken for missing data. Given the size of the study data set, no meaningful loss of power was anticipated due to missing data. The level of imbalance in the complete data set from the BSA meeting the study inclusion criteria was ascertained through visual inspection and consideration of summary statistics for the variables on which the matching was based. We also reported summary statistics for the outcome variables for the unmatched data set and the matched data set, overall and by exposure status. The software packages R130 and Stata 14131 were used for all statistical analysis.

The R package MatchIt108,132 was used to estimate the propensity scores and undertake the matching of non-exposed individuals to exposed individuals. The propensity scores were estimated using logistic regression, where the response variable was exposure to water fluoridation taking a value of 0 for non-exposed (grand mean fluoride concentration of < 0.7 between 2010 and 2020, and a value of 1 for exposed). The model covariates were those identified by the methods described earlier. We planned to estimate the propensity scores initially in a parsimonious model using main effect terms only and restricting the inclusion of higher order and interaction terms unless adequate balance could not be achieved. This approach was undertaken to minimise the risk of overfitting. 106

Distance matching was used to match exposed and non-exposed individuals. We proposed to employ three methods of distance matching: nearest neighbour, optimal and full matching, with exact matching on the covariates sex and area level deprivation. The use of multiple matching algorithms has been recommended to select the match that has achieved ‘greatest’ covariate balance (where the distribution of the covariates is similar across the exposed and non-exposed groups) while maintaining an adequate sample size through minimisation of discarded observations. 133

Initially we planned to use a nearest neighbour matching algorithm. This is the simplest form of matching and involves finding and matching individuals in un-exposed group who have the closest propensity scores to those in the exposed group. 133 This method has the advantage that it works well for estimating the average treatment effect in the treated (ATT) (i.e. exposed to water fluoridation) when the number of treated cases is substantially smaller than the number of untreated cases available for matching, as is the case for our data set. The ATT in this study is the average effect that would have been expected if everyone in the exposed group received water fluoridation at the 0.7 threshold compared to if no one in this group received the exposure. 106 One limitation to the nearest neighbour matching approach, however, is that the best available match is sought without accounting for the resulting quality of the match of the entire treated sample.

We planned to refine the matching specification by considering the use of a calliper based on the propensity score, and by using a variable ratio matching as well as fixed one to one matching. A calliper is a maximum permissible distance in propensity scores for the match. It can be defined as ‘a fraction of the standard deviation of the logit of the propensity score’. 133 To remove at least 90% of bias, Rosenbaum and Rubin (1985) used a calliper of 0.25 SDs. 134 Any individuals whose propensity scores are not within the specified calliper (e.g. 0.25 SDs) of an individual in the comparison group are not selected for matching. The use of a calliper therefore increases the quality of the matches but risks reducing the matched sample size.

We also planned to undertake optimal pair matching and optimal full matching, which, in contrast to nearest neighbour matching, attempt to optimise the quality of the match for the overall matched data set, but these methods do require substantial processing power for the optimised matching of large data sets. Further description of these matching methods can be found in the publication by Lee and Little (2017). 106

Our proposed matching methods were:

• Method 0: No matching

• Method 1: Nearest neighbour, one to one matching

• Method 2: Nearest neighbour, variable ratio matching (ratio = 3, maximum = 5)

• Method 3: Nearest neighbour, variable ratio matching (ratio = 3, maximum = 5), with 0.25 calliper

• Method 4: Optimal matching

• Method 5: Optimal matching, variable ratio matching (ratio = 3, maximum = 5)

• Method 6: Full matching

• Method 7: Full matching with calliper.

Nearest neighbour matching with a fixed one-to-one match is the simplest model and was chosen as our initial approach, given the size of the data set and the number of exposed observations for whom we are seeking a match. An important limitation however is this method results in the smallest matched data set as potentially suitable non-exposed individuals are discarded once one non-exposed individual has been matched. Consequently, the total matched data set can only ever be as large as twice the size of the exposed group (but may be smaller).

This shortcoming was addressed with methods 2 and 3 using variable ratio matching. With variable ratio matching the number of matched non-exposed matches is permitted to vary, thus limiting the number of unsuitable matches compared to a fixed ratio approach. A larger sample size with more similar matches is thereby retained. As stated in Leite,133 research has shown variable ratio matching removes more bias than one-to-one matching135–137 and is particularly useful if the number of untreated (un-exposed) group cases is much larger than the number of treatment cases138 as is the case with the exposed and non-exposed groups in the study.

Method 2 was further refined in method 3, where a 0.25-unit calliper (a maximum permissible distance in propensity scores for the match) was added to the variable matching. In fixed and variable matching, the chance of unbalanced covariates may remain when the additional matches are not always as close as the first match. 108,132

In order to determine the ‘success’ of the different matching algorithms and to select the matched data set for analysis we considered the number of matched participants, including the number and impact of discarded observations from the exposed group, the overall sample size and the distribution of the covariates across the exposed and non-exposed groups. Evaluation of covariate balance was undertaken through visual inspection of their distributions in the unexposed and exposed groups, and examination of descriptive statistics. Visual inspection of covariate balance was undertaken through Love plots, empirical quantile–quantile (eQQ) plots for continuous covariates, bar charts for categorical covariates and kernel density plots for propensity score distribution and continuous variables. Balance was deemed to have been achieved on the eQQ plots where values fall on the 45^o line, and for the empirical cumulative distribution function (eCDF) and kernel density plots where the lines of the exposed and non-exposed groups overlap.

Given the large sample size of the study data set, standardised mean differences (SMDs) were used in preference to hypothesis tests to determine covariate balance. 106 The SMD for a given covariate compares the difference in the means of the covariate (between the exposed and unexposed groups), divided by the pooled SD of the covariate. 139 SMD is also referred to as Cohen’s d,140 often used to frame effect sizes as small, medium or large in the psychological literature. 141 The large sample size meant it was conceivable that statistically significant differences in SMDs obtained through hypothesis testing for the covariates would be indicative of unimportant mean differences. In preference, absolute SMDs below a threshold of 0.1 SDs were considered to constitute adequate balance. 142

Lastly, we compared the distributions of the propensity scores for the exposed and non-exposed groups to ascertain the existence of common support for the treated observations through visual inspection of the linear propensity scores for the treated and untreated (jitter plots and histograms). If a substantial overlap is observed over the range of the propensity scores between the treatment and comparison groups, then this suggests a large area of common support over the participants. 106

Ancillary analyses

1. An additional analysis was proposed by the Study Oversight Committee. We modelled the probability of receiving an invasive dental treatment according to exposure using logistic regression.

2. An offset term using the total number of courses of treatment was added to the regression model for the primary outcome of number of invasive dental treatments, so the exposure of the optimally fluoridated and sub/non-optimally fluoridated groups is proportional to the number of courses of treatment. Including an offset term in the model adjusts for the opportunity that an event has to occur. To determine whether there was a differential effect of water fluoridation according to deprivation on the primary outcome, when taking account of the number of courses of treatment received, we undertook an additional regression analysis that included the main effects of water fluoridation and deprivation decile with an interaction term and included an offset term for total number of courses of treatment.

Approach

The health economics analysis consisted of calculating the ROI and a cost-effectiveness analysis. For these analyses, a public sector perspective was taken, which accounted for the direct costs and savings accrued to the relevant public sector bodies. The costs of water fluoridation are borne by PHE (for capital infrastructure costs), or local authorities (for operating costs), with any savings in treatment costs accrued by the NHS. Indirect public sector and/or wider societal costs were not included due to the pragmatic nature of the study and the limitations in the available data.

The ROI analysis allows for the financial costs for water fluoridation to be compared against the financial returns resulting from the NHS paying for fewer primary care dental treatments (all types of treatment). The cost-effectiveness analysis compares these same costs against avoiding invasive dental treatments; these being irreversible surgical procedures that often provoke anxiety and result in the loss of healthy tooth tissue.

Data collection

Cost effectiveness

Cost effectiveness was based on the primary study outcome, assessed as the additional cost per avoided episode of invasive dental treatment. A public sector perspective was taken, as the costs were paid by PHE and the effectiveness is measured at NHS dental practices, and we estimated the incremental cost-effectiveness ratio (ICER). The ICER was measured by the difference in water fluoridation costs between optimally fluoridated and sub/non-optimally fluoridated groups (incremental costs), divided by the difference in the number of invasive dental treatments between optimally fluoridated or sub/non-optimally fluoridated groups (incremental effects).

The per-patient difference in the number of invasive dental treatments was provided by the statistical analysis detailed in Data validation and cleaning. Per-person differences estimated from our sample were extrapolated to the relevant population of regular users of NHS dental services living in optimally fluoridated areas using the same scenarios detailed above.

Reductions in dental service costs were not factored into the incremental costs of fluoridation (numerator in the ICER) because the reduction in invasive dental treatments is used as the measure of effects (denominator in the ICER). Deducting the costs of reduced dental treatments from the costs of water fluoridation would involve ‘double counting’ this change in treatment episodes.

The time horizon was the study duration. Costs and outcomes were not discounted as the difference in outcomes is observed once, at the end of study follow-up period and therefore the year in which differences occurred was not known.

Sensitivity analysis using fee-for-item costs

To afford greater sensitivity to the examination of relative differences in the use of NHS resources by patients in the optimally fluoridated and sub/non-optimally fluoridated groups, an alternative costing approach informed by the Scottish NHS fee-per-item system was also utilised as a sensitivity analysis. This is because in the English NHS dental contract, a course of treatment containing multiple invasive dental treatments attracts the same NHS payment (UDA allocation) as a course of treatment containing just one invasive dental treatment. It could be argued that the UDA system (with just three ‘bands’ of costs) could obscure cost differences as a result of lower treatment needs. Investigating the effect of using the Scottish fee-for-item costs was hypothesised to provide a more discriminatory approach.

It is important to note the intention here is not to estimate ‘equivalent’ Scottish NHS treatment costs, and the values attached to each treatment item are not intended to accurately quantify resource use in absolute terms. It is included for comparison purposes of relative differences only. This approach to costing dental care may also be more understandable to readers who are not familiar with the English system of remuneration for NHS dentistry.

For each course of treatment, the NHS BSA clinical data set contained the total number of treatments provided within each of the following treatment categories: examinations; permanent fillings and sealants; extractions; inlays; crowns; veneers; endodontic treatment and fluoride varnish.

The fee-for-item costs for each type of treatment in the Scottish Dental Remuneration schedule 142 November 2019153 were reviewed by two dentists who have practised in the NHS system in England and one dentist who has practised in the NHS system in Scotland. The Scottish treatment item that would best reflect the ‘middle’ level of complexity for each type of treatment recorded in the English FP17 clinical data set was chosen as the reference cost. The total number of treatments in each category for each patient was then multiplied by the chosen reference cost.

Substantial amendment 1 (NHS REC approved 17th of September 2020)

In the original proposal, the ‘fluoridated’ group were to be defined as individuals living in an area with an artificial water fluoridation programme (regardless of achieved water fluoride concentrations). The effect of achieved water fluoride concentrations were to be considered in a subgroup analysis investigating the differential effects of receiving optimally (≥ 0.7 mg F/l), or suboptimally fluoridated water (< 0.7 mg F/l). However, it was not possible to source an indicator of which areas in England were part of a fluoridation programme; therefore, we were unable to carry out the analysis as planned. The method for group assignment was changed to be based solely on measured water fluoride concentrations rather than programme coverage.

Other changes

In the protocol, the wording of the age range of the participants is stated as ‘adolescents and adults aged over 12 years’. We wish to clarify that the intended meaning should read as ‘adolescents and adults aged 12 years and over’ (i.e. beyond their 12th birthday).

For the secondary objective related to oral health, we had originally planned to use as ‘number of remaining teeth’ as an outcome, in addition to DMFT. The data are provided by NHS BSA as individual D, M and F components. During the analysis, we decided to avoid the planned transformation of 32-missing and to use the data as supplied by NHS BSA number of missing teeth. This was because the raw data had a better fitting distribution for the statistical models, and we could not be sure of the original number of teeth in the subtraction (32 vs. 28), as many people have congenitally missing third molars (wisdom teeth).

On 28 January 2022, the study end date changed to reflect a 6-month extension as a result of COVID-19 and the additional work involved in sourcing and compiling water fluoride concentration data.

Resulting matched data sets

• Method 0 The unmatched data set consisted of 17,796,185 observations, with 1,592,570 individuals ascribed optimally fluoridated status and 16,203,615 individuals ascribed sub/non-optimally fluoridated status.

• Method 1 (nearest neighbour, one to one matching): The use of nearest neighbour matching with a one-to-one match of controls (sub/non-optimally fluoridated) to treated individuals (optimally fluoridated) resulted in the smallest sample size of the nearest neighbour methods. Matches were found for all 159,270 exposed individuals and common support was found for all treated cases.

• Method 2 (nearest neighbour, variable ratio matching): The use of nearest neighbour matching with a one-to-three variable ratio match of controls (sub/non-optimally fluoridated) to treated individuals (optimally fluoridated) resulted in a total sample size of 6,370,280, the largest of all nearest neighbour matches. Matches were found for all 159,270 exposed individuals and common support was found for all treated observations. No exposed observations were discarded.

• Method 3 (nearest neighbour, variable ratio matching, calliper): The use of nearest neighbour with variable ratio matching and the addition of a 0.25 calliper resulted in a total sample size of 6,369,406. Four hundred and thirty-seven observations from the intervention (optimally fluoridated) group could not be matched within this calliper and were subsequently discarded from the matched data set. Consequently, the estimand using this method no longer corresponds to the ATT (exposed), but rather an undefined matched population.

The sample sizes for the three matching methods are reported in Table 4. The total sample size, with the numbers of matched and unmatched (discarded) individuals for the exposed and non-exposed groups are presented.

Evaluation of covariate balance for methods 1 through 3 was undertaken through visual inspection and examination of descriptive statistics. Visual inspection of covariate balance was undertaken through Love plots, eQQ and kernel density plots for continuous covariates, and bar charts for categorical covariates. Balance was deemed to have been achieved on the eQQ plots where values fall on the 45° line, and for the kernel density plots where the lines of the exposed and non-exposed groups overlap. Additionally, values of absolute SMDs were used in preference to hypothesis tests to determine covariate balance, with SMDs below a threshold of 0.1 SDs considered to constitute adequate balance. Based on all the above information Model 2 was chosen to take forwards for further analysis.

Results indicated that methods 1 and 2 had worked well to produce a matched data set, convincingly meeting the threshold values for absolute SMD and used all available information from the exposed observations. With method 3, 437 treated cases were discarded from the matched data set, while the observed percentage balance improvement over the unmatched data set was the same as for method 2 (99.9%). Consequently, the matched data set resulting from method 3 was not taken forwards for analysis. Based on the larger sample size and the fact that variable ratio matching has been shown to remove more bias than one-to-one matching, we elected to take forwards method 2 as the matched data set for analysis.

Balance of the selected matched cohort

A Love plot showing the SMDs for the unmatched and matched analytical cohorts is shown in Figure 10. The dashed vertical lines in the Love plot represent a SMD of 0.1 SDs (the recommended threshold for adequate covariate balance without further regression-based adjustments). The solid vertical line represents a SMD of zero. The red indicators are the SMDs in the unmatched cohort and the blue indicators represent the SMDs in the selected matched cohort (method 2). All absolute standardised effect sizes (SMDs) of the matched sample were below the threshold of 0.1 SDs and were therefore considered to constitute adequate balance. Additional graphs evaluating covariate balance for method 2 are presented in Appendix 5, Figures 22–33. Table 5 presents summary balance statistics for the selected matched analytical cohort.

FIGURE 10.

Love plot illustrating SMD for the unmatched cohort and the matched cohort taken forwards for analysis (method 2).

As shown in Table 5 the SMDs between the exposed and unexposed group were all below the recommended threshold for adequate covariate balance of 0.1. 142 The SMDs for the patient deprivation and sex variables are zero because exact matching was used for these variables. The per cent balance improvement is also shown which indicates how improved the SMD was by the matching procedure (compared to the unmatched data set).

Table 6 illustrates the variance ratio (VR) for the continuous variables of age and UDAs commissioned per person (local authority level). VR is the ratio of the variance of a covariate in one group to the variance of the same covariate in the comparator group. VRs close to one are considered to indicate good balance. 141

Baseline characteristics of the unmatched and matched cohorts

Table 7 presents the baseline individual-level characteristics of the unmatched (full) study cohort and the selected matched study cohort, by fluoridation group. Table 8 presents the baseline contextual characteristics for the same cohorts. The individual and contextual baseline characteristics presented are those which the propensity score matching was based on. All individual level and contextual baseline characteristics were derived from information recorded at the patient’s first course of dental treatment in the study observation period.

Clinical characteristics and health outcomes for the unmatched and matched cohorts

Table 9 presents descriptive summary statistics for the clinical characteristics of the unmatched and matched cohorts overall and by water fluoridation status. The clinical characteristics of the matched and unmatched cohorts are relatively similar. This provides reassurance the findings from the matched cohort can be generalised to the wider population of NHS dental service users.

We carried out an analysis of DMFT and missing teeth by age group, to investigate the possibility that differences in the number of missing teeth may be more sensitive to the effects of water fluoridation in older people, whereas decayed and filled teeth may be more sensitive in younger people. This was informed by recent studies from the UK and Australia which suggest this may be the case. 6,66,80 Age 45 was chosen as the cut-point because this is the age at which differences have been reported previously. 66

These analyses were carried out on the full unmatched cohort because it would not be possible to split the matched cohort by age without splitting the matched subclasses.

Tables 10 and 11 show that the mean DMFT and mean number of missing teeth are very similar between the optimally fluoridated and sub/non-optimally fluoridated groups, in both the under 45 and 45 and over age groups.

Primary outcome: invasive dental treatments

The primary research question of interest was whether water fluoridation reduced the number of invasive dental treatments (restorations, endodontics, extractions) received by adults attending NHS dental practices over 10 years of observation (22 November 2010 to 21 October 2020).

Outcome information was available for 6,370,280 observations in total, with 4,777,710 observations in the sub/non-optimally fluoridated group and 1,592,570 in the optimally fluoridated group. The mean number of invasive treatments per person in the matched study sample overall was 5.573 treatments (SE 0.003), with a similar number of mean treatments per person in the sub/non-optimally fluoridated (5.616, SE 0.005) and optimally fluoridated groups (5.443, SE 0.005). The median number of invasive treatments was four, with a maximum number of invasive dental treatments of 271 in a distribution that was positively skewed (Figure 12 and Table 12). Overall, 84.5% patients in the matched sample received at least one invasive dental treatment in the 10-year period of observation.

FIGURE 12.

Histogram illustrating the distribution of the total number of invasive dental treatments per patient.

As the outcome is a count of the total number of invasive treatments received, we used a negative binomial regression to compare the groups and determine the effect estimate, accounting for the clustering of the data through cluster robust SE and weighting the data. A negative binomial regression was chosen rather than a Poisson regression to account for the substantial overdispersion in the distribution for this outcome.

The IRR was 0.969 (95% CI 0.967 to 0.971), indicating the rate of invasive treatments in the optimally fluoridated group was 3% less than that of the sub/non-fluoridated group. The predicted count of the number of invasive treatments from this model was 5.616 (SE 0.004) for the sub/non-optimally fluoridated group and 5.443 (SE 0.005) for the optimally fluoridated group. There was a predicted mean difference of 0.173 fewer invasive dental treatments per person (95% CI 0.185 fewer to 0.161 fewer) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group.

To aid with readers with the interpretation of the relevance of any differences in the primary outcome, we carried out stakeholder engagement to identify what would be considered to be a meaningful ‘per person’ reduction in the number of invasive dental treatments (for everyone in a hypothetical town), as a result of water fluoridation. In the engagement workshops we presented a scenario that six fillings or extractions would be received in the sub/non-optimally fluoridated group over 10 years and sought views on what size of reduction would be considered meaningful. The scenario presented during the workshops was in fact very similar to what we found, with the mean number of invasive dental treatments received in the sub/non-optimally fluoridated group in our matched cohort being 5.6. This makes the comparison of the effect size for our primary outcome against stakeholder views fairly straightforward with regard to relative and absolute differences.

During the engagement workshops, most participants felt that an average reduction of at least one invasive NHS dental treatment per person over 10 years would be needed to be considered as meaningful to them (a 16% relative reduction in the scenario). The dentists present did consider smaller reductions of 0.30 fewer invasive dental treatments per person to be meaningful (a 5% relative reduction in the example scenario), but only a minority of all stakeholders (4 out of 54) felt the smallest reduction presented in the scenarios (0.18 fewer treatments, a relative difference of 3%) would be considered meaningful. Effects sizes of 3% relative difference or smaller have previously been referred to as ‘exceedingly small’ in the epidemiological literature. 161

Secondary outcomes

The secondary outcomes of interest in this study were the number of DMFT and the number of missing teeth.

Primary outcome: invasive dental treatments

The summary estimates for number of invasive dental treatments are presented in Table 13 by water fluoridation status and deprivation decile. There was a clear pattern to the mean number of invasive dental treatments according to area-level deprivation for this outcome, overall and by group. From the most deprived groups (IMD decile 1) to the least deprived groups (IMD decile 10), the mean number of invasive treatments received decreases. This pattern is evident in both the optimally fluoridated and sub/non-optimally fluoridated groups.

TABLE 13 - Summary estimates for mean number of invasive dental treatments received, by water fluoridation status and deprivation decile

Deprivation decile (IMD)	Sub/non-fluoridated N, mean (SE)	Optimally fluoridated N, mean (SE)	Total N, mean (SE)
1 (most deprived)	813,723 6.324 (0.012)	271,241 5.988 (0.012)	1,084,964 6.241 (0.010)
2	685,104 6.056 (0.012)	228,368 5.788 (0.013)	913,472 5.989 (0.010)
3	493,686 5.756 (0.013)	164,562 5.675 (0.015)	658,248 5.736 (0.010)
4	413,940 5.565 (0.012)	137,980 5.455 (0.016)	551,920 5.537 (0.010)
5	425,658 5.459 (0.012)	141,886 5.278 (0.016)	567,544 5.414 (0.010)
6	392,451 5.357 (0.011)	130,817 5.171 (0.016)	523,268 5.310 (0.009)
7	406,113 5.286 (0.011)	135,371 5.175 (0.016)	541,484 5.259 (0.009)
8	408,306 5.159 (0.011)	136,102 5.097 (0.016)	544,408 5.143 (0.009)
9	394,791 5.057 (0.011)	131,597 5.001 (0.016)	526,388 5.043 (0.009)
10 (least deprived)	343,938 4.985 (0.012)	114,646 4.867 (0.017)	458,584 4.956 (0.010)
Total	4,777,710 5.616 (0.004)	1,592,570 5.443 (0.005)	6,370,280 5.573 (0.003)

We modelled the effects of water fluoridation and area-level deprivation on the number of invasive dental treatments using a negative binomial regression model with main and interaction effects, cluster robust SE and accounting for weighting. The number of observations for this outcome was 6,370,280.

The conditional effect estimate for the rate of invasive dental treatments according to water fluoridation status from the model with the interaction term was IRR 0.976 (95% CI 0.968 to 0.984). The predicted marginal mean values for the number of invasive dental treatments were 5.443 (SE 0.005) in the optimally fluoridated group and 5.616 (SE 0.004) in the sub/non-optimally fluoridated group. The effect estimates from the regression model for the main and interaction terms along with their CIs are reported in Appendix 6, Tables 38–40

Figure 12 illustrates the mean predicted number of invasive dental treatments for the optimally fluoridated and sub/non-optimally fluoridated groups at each deprivation decile, from the negative binomial regression model. The mean predicted number of invasive dental treatments for the matched sample overall is highest for the most deprived decile (IMD decile 1), decreasing steadily to the least deprived decile (IMD decile 10). Compared to the effect of water fluoridation in deprivation decile 10, the reference category of the least deprived, the effect of water fluoridation is greatest in most deprived deciles 1 and 2.

Table 14 shows the discrete changes in the predicted number of invasive treatments (the difference in the predicted values of the mean number of invasive treatments in the optimally fluoridated group and the mean number in the sub/non-optimally fluoridated group) at each deprivation decile. In IMD decile 1 (most deprived), the predicted mean number of invasive dental treatments received is 0.337 lower (95% CI −0.371 to −0.302) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group. In IMD decile 2, the predicted mean number of invasive dental treatments received is 0.268 lower (95% CI −0.304 to −0.233) lower in the optimally fluoridated group compared to the sub/non-optimally fluoridated group (Figure 13 and Table 11). In the remaining deciles the differences are smaller. The effects of water fluoridation do not bring the mean number of predicted invasive dental treatments in the more deprived deciles in line with those of the least deprived deciles.

TABLE 14 - Difference in the negative binomial predicted values of the mean number of invasive treatments, comparing the optimally fluoridated group to the sub/non-optimally fluoridated group at each deprivation decile

Deprivation decile (IMD)	Difference in the predicted mean number of invasive dental treatments (95% CI) (negative values favour water fluoridation)	p > |z|
1 (most deprived)	−0.337 (−0.371 to −0.302)	0.000
2	−0.268 (−0.304 to −0.233)	0.000
3	−0.080 (−0.119 to −0.042)	0.000
4	−0.110 (−0.149 to −0.070)	0.000
5	−0.181 (−0.220 to −0.143)	0.000
6	−0.186 (−0.224 to −0.147)	0.000
7	−0.111 (−0.149 to −0.074)	0.000
8	−0.062 (−0.099 to −0.025)	0.001
9	−0.055 (−0.093 to −0.018)	0.004
10 (least deprived)	−0.118 (−0.158 to −0.078)	0.000

FIGURE 13.

Predicted number of invasive dental treatments with 95% CIs by water fluoridation status and patient deprivation decile from negative binomial regression model.

The predicted effects of optimal water fluoridation on the number of invasive dental treatments received within each decile would not have been considered meaningful by most stakeholders that we engaged with, though the dentists we consulted did consider differences as small as 0.3 fewer invasive dental treatments to be meaningful.

Secondary outcome: DMFT

The summary estimates for DMFT are presented in Table 15 by water fluoridation status and deprivation decile for the subset of the matched sample post 2016. There was no clear association of DMFT count with area-level deprivation overall or by water fluoridation status.

TABLE 15 - Summary estimates for DMFT by water fluoridation status and deprivation decile

Deprivation decile (IMD)	Sub/non-optimally fluoridated (n = 4,097,227) N, mean (SE)	Optimally fluoridated (n = 1,419,879) N, mean (SE)	Total (n = 5,517,106) N, mean (SE)
1 (most deprived)	349,879 13.009 (0.017)	250,347 12.024 (0.018)	600,226 12.763 (0.014)
2	399,225 13.229 (0.018)	207,813 12.804 (0.019)	607,038 13.123 (0.015)
3	396,660 13.189 (0.019)	146,300 13.262 (0.022)	542,960 13.207 (0.016)
4	418,761 13.182 (0.019)	120,457 13.315 (0.025)	539,218 13.216 (0.016)
5	403,014 13.331 (0.019)	125,162 13.506 (0.024)	528,176 13.375 (0.016)
6	437,699 13.606 (0.018)	113,687 13.616 (0.026)	551,386 13.608 (0.015)
7	439,602 13.654 (0.018)	118,650 13.722 (0.025)	558,252 13.671 (0.015)
8	443,079 13.595 (0.018)	119,691 13.774 (0.025)	562,770 13.639 (0.015)
9	434,143 13.634 (0.018)	115,859 13.637 (0.025)	550,002 13.635 (0.015)
10 (least deprived)	375,165 13.789 (0.019)	101,913 13.346 (0.027)	477,078 13.676 (0.017)
Total	4,097,227 13.361 (0.006)	1,419,879 13.149 (0.007)	5,517,106 13.308 (0.005)

We modelled the effects of water fluoridation and area-level deprivation on DMFT using a negative binomial model with main and interaction effects, cluster robust SE and accounting for weighting. The number of observations for this outcome was 5,517,106. The conditional effect estimate for the effect of water fluoridation was IRR 0.968 (95% CI 0.964 to 0.973), with predicted mean values for the DMFT count of 13.149 (SE 0.007) in the optimally fluoridated group and 13.361 (SE 0.006) in the sub/non-optimally fluoridated group. The effect estimates for the main and interaction terms along with their CIs are reported in Appendix 6, Tables 38–40.

Figure 13 illustrates the predicted mean DMFT count for the optimally fluoridated and sub/non-optimally fluoridated groups at each deprivation decile from the negative binomial model. Table 16 shows the discrete changes in the predicted mean DMFT count (the difference in the predicted values of the mean DMFT count in the optimally fluoridated group and the mean number in the sub/non-optimally fluoridated group) at each deprivation decile.

TABLE 16 - Difference in the negative binomial predicted values of mean DMFT count, comparing the optimally fluoridated group to the sub/non-optimally fluoridated group at each deprivation decile

Deprivation decile (IMD)	Difference in the predicted mean DMFT count (95% CI) (negative values favour water fluoridation)	p > |z|
1 (most deprived)	−0.986 (−1.033 to −0.939)	0.000
2	−0.426 (−0.476 to −0.375)	0.000
3	0.073 (0.016 to 0.13)	0.012
4	0.132 (0.072 to 0.192)	0.000
5	0.174 (0.116 to 0.233)	0.000
6	0.01 (−0.049 to 0.069)	0.740
7	0.068 (0.01 to 0.126)	0.022
8	0.179 (0.121 to 0.236)	0.000
9	0.004 (−0.054 to 0.062)	0.894
10 (least deprived)	−0.440 (−0.501 to −0.378)	0.000

Unlike the analyses for the primary outcome, where a social gradient in the expected direction from most to least deprived was evident, there was no such pattern for DMFT. In IMD decile 1, the predicted mean DMFT count is 0.986 lower (95% CI 1.033 lower to 0.939 lower) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group. In IMD decile 2, the predicted mean DMFT count is 0.426 lower (95% CI 0.476 lower to 0.375 lower) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group (Figure 14 and Table 16). However, the direction of effect is not consistent across the deciles. In IMD deciles 3–9, the mean predicted DMFT count is either very similar in both groups (IMD deciles 6, 7 and 9) or is slightly higher in the fluoridated group than in the sub/non-optimally fluoridated group (IMD deciles 3, 4, 5 and 8).

FIGURE 14.

Predicted number of DMFT with 95% CIs by water deprivation status and deprivation decile from negative binomial regression model.

Secondary outcome: number of missing teeth

The summary estimates for the number of missing teeth are presented in Table 17 by water fluoridation and deprivation decile. Similar to the mean number of invasive dental treatments but different to DMFT, there was evidence of a clear pattern according to area-level deprivation on this outcome for the subset of the matched sample post 2016 overall and by water fluoridation status. As level of deprivation decreases, from the most deprived (IMD decile 1) to the least deprived groups (IMD decile 10), the mean number of missing teeth decreases, and this pattern is evident in the sub/non-optimally fluoridated group and for most deciles of the optimally fluoridated group.

We modelled the effects of optimal water fluoridation and area level deprivation on the number of missing teeth using a negative binomial model with main and interaction effects, cluster robust SE and accounting for weighting. The number of observations for this outcome was 5,517,106. The conditional effect estimate for the effect of optimal water fluoridation was IRR 0.984 (95% CI 0.977 to 0.992), with predicted mean values for the number of missing teeth of 6.652 (SE 0.006) in the optimally fluoridated group and 6.646 (SE 0.005) in the sub/non-optimally fluoridated group. The effect estimates for the main and interaction terms along with their CIs are reported in Appendix 6.

Figure 14 illustrates the predicted mean number of missing teeth for the optimally fluoridated and sub/non-optimally fluoridated groups at each deprivation decile. Table 18 shows the discrete changes in the predicted mean number of missing teeth (the difference in the predicted values of the mean number of missing teeth in the fluoridated group and the mean number in the sub/non-optimally fluoridated group) at each deprivation decile.

In IMD decile 1, the predicted mean number of missing teeth is 0.476 lower (95% CI 0.517 lower to 0.435 lower) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group. In IMD decile 2, the predicted mean number of missing teeth is 0.174 lower (95% CI 0.217 lower to 0.130 lower) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group (Figure 15 and Table 18). However, as with DMFT, the direction of effect is not consistent across the deciles. In IMD deciles 3–9, the predicted mean number of missing teeth is higher in the optimally fluoridated group than in the sub/non-optimally fluoridated group; for example, in IMD decile 8 the optimally fluoridated group is predicted to have 0.325 more missing teeth per person than the sub/non-optimally fluoridated group.

FIGURE 15.

Predicted mean number of missing teeth with 95% CIs by water fluoridation status and deprivation decile from negative binomial regression.

Ancillary analysis

Water fluoridation costs

Data on water fluoridation expenditure (provided by PHE) are summarised in Table 19 alongside the estimated population receiving artificially fluoridated water (≥ 0.7 mg F/l) and per person expenditure. The population receiving artificially fluoridated water was estimated by combining the water fluoridation sample data and maps used to assign patient to optimally fluoridated or sub/non-optimally fluoridated groups (see Water fluoridation), with published estimates and maps detailing the extent of natural and artificial water fluoridation in England. 21,151

Capital expenditure includes replacement plants or repairs and upgrades to fluoride dosing equipment and facilities and is subject to annual fluctuations based on the required work. Operating expenditure included running costs such as workforce, transport and chemicals. Expenditure data could not be provided by PHE for the years 2010–2 as this is a period before PHE existed. For these years the mean expenditure was used, calculated over the period 2013–9. The estimated size of the population receiving artificial fluoridation at the required dose was used to calculate expenditure per person. The GDP deflator is used to express these costs in 2020 prices. Total expenditure between 2010 and 2019 was £46,791,388 in 2020 prices. The average size of the artificially optimally fluoridated population between 2010 and 2019 was 4,544,172 yielding a per person optimal fluoridation cost over 10 years of £10.30 in 2020 prices.

Dental resource use and costs

The sample consists of 6,370,280 patients who are split between 1,592,570 in the optimally fluoridated group and 4,777,710 in the sub/non-optimally fluoridated group. Table 20 summarises the total courses of treatment per patient from 2010 to 2020. Mean total courses of treatment were 0.22 lower in the optimally fluoridated group, 9.91 versus 10.13. This difference is driven by very high service uses and the percentiles in the distribution are the same until the 99th percentile.

Table 21 shows how treatment bands are distributed between those in the optimally fluoridated and sub/non-optimally fluoridated groups. Optimally fluoridated patients had a greater percentage of band 1 treatments and slightly lower percentage of band 2 and band 3 treatments. Urgent treatments were also lower for optimally fluoridated patients.

Table 22 shows NHS patient changes per patient for the period 2010–20. This includes patients who qualified for NHS charge exemptions at some/all of their courses of treatments. Optimally fluoridated patients had lower patient charges, but the mean difference is relatively small at £7.64. From the 75th percentile the differences become greater, and at the 99th percentile the difference is £50.

Dental practices are reimbursed for the care provided using UDAs and the financial value of a UDA is negotiated. There is considerable variation in the financial value of 1 UDA between contracts. Our analysis uses the mean cost per UDA for all practices in England between financial years 2015–6 and 2019–20, obtained from NHS BSA via a freedom of information (FOI) request (Table 23). Treatment bands have different UDA multipliers: Band 1 × 1 UDA, Band 2 × 3 UDA, Band 3 × 12 UDA and Urgent × 1.2 UDA.

The average cost per UDA and treatment band multiplier are used to generate total UDA treatment costs shown in Table 24. Each course of treatment is assigned the UDA cost paid to the practice by the NHS and then totalled for each patient from 2010 to 2020. Patients in fluoridated areas had a lower UDA treatment cost at £755.75 versus £784.34, a difference of £28.59. As with patient charges, the difference increases for those with higher service use and reaches £191.63 at the 99th percentile.

For each course of treatment, the NHS pays the dental practice the UDA treatment costs but also recovers the patient charge amount, where these charges are applicable. Table 25 presents the NHS treatment cost, where patient charges have been deducted from UDA treatment costs to give the total cost to the NHS. Patients in optimally fluoridated areas had a lower NHS treatment cost at £385.75 versus £407.70, a difference of £21.95. This difference increases to £135.96 for the 99th percentiles.

The difference in NHS treatment costs between optimally fluoridated and sub/non-optimally fluoridated patients is estimated using a GLM and the regression results are shown in Table 26. The marginal effects estimate reveals a saving in treatment costs for optimally fluoridated patients over the study period of £22.26 per person with a narrow 95% CI of £21.43 to £23.09. This represents a relative reduction in NHS treatment costs of 5.46% per person (95% CI −5.26% to −5.66%) in the optimally fluoridated group.

Return on investment

The return on investing in water fluoridation is estimated using the cost provided by PHE and the estimated NHS cost saving from Table 20. As the cost saving is estimated on a sample, we present four alternative scenarios to estimate the proportion of the optimally fluoridated population these cost savings can be attributed.

1. Scenario 1: estimated cost savings are attributed to the average optimally fluoridated population size 2010–20. This scenario will overestimate the ROI as it assumes our sample and the fluoridated population have the same characteristics which we know is not the case. For example, our sample is aged 12 + and have received at least two courses of NHS dental treatment in the study period.

2. Scenario 2: estimated cost savings are attributed to an estimate of the proportion of patients who have received at least two courses of NHS dental treatment in the study period. The proportion is estimated from responses to the GP Patient Survey 2019. Respondents were asked for reasons for not having an appointment in the last 2 years and could answer: prefer private; stayed with dentist when changed from private; tried but not successful; cannot remember if successful or not; and didn’t think they could get one. The proportion giving these reasons for not having an appointment summed to 24.40%; hence we estimate the proportion who may have received at least two courses of NHS dental treatment within 10 years to be 75.60%.

3. Scenario 3: estimated cost savings are attributed to the proportion of patients who have seen an NHS dentist within the NICE maximum recommended recall period, as reported by NHS BSA. The NICE recommended recall period is within 12 months for those aged 12–18 years and within 24 months for 18 +. This will be the lower boundary of estimate of the true proportion because the time frames are shorter than the 10 years. This proportion is 50.30%.

4. Scenario 4: estimated cost savings are attributed to a proportion of patients lying equally between scenario 2 and 3. This scenario was suggested by members on this study steering group and represents what they believed to be closest to the actual figure to an estimate of the proportion of the 12 + population who would receive at least two courses of NHS dental treatment in 10 years. The proportion is 62.95%.

The ROI for each scenario is shown in Table 27.

Taking scenario 4 as an example, we see the investment in fluoridation costing £46.8 million and reducing NHS treatment cost by £63.7 million over the study period. This yields a ROI of £16.9 million, representing 36% of the total investment. In this calculation, the only benefit is achieved via reducing NHS treatment costs for the scenario population. No other plausible benefits are factored, which could include lower treatment costs outside of the population (i.e. those aged under 12) or benefits other than NHS treatment costs such as reduced school or work absences.

Cost effectiveness

The cost effectiveness of water fluoridation is estimated using the cost provided by PHE and the estimated number of invasive treatments avoided from Clinical results (the primary outcome).

As with the ROI, four scenarios are used and calculations for cost effectiveness are shown in Table 28.

The incremental costs are £46.8 million as the fluoridation costs are zero in the sub/non-optimally fluoridated group. The incremental benefits vary depending on the scenario chosen, taking scenario 4 as an example, we see the number of invasive treatments decreasing by 494,876. This leads to a cost of £94.55 per invasive treatment avoided. ICER=£46.8m−£05.616−5.443=£94.55

This ICER of £94.55 cannot alone be used to determine if the water fluoridation scheme were cost-effective or not, when evaluated using the study period and study population. For that determination a threshold value would be needed, a monetary value below which the ICER would represent a cost-effective decision. It should be noted that the benefit of avoiding an invasive dental treatment cannot be solely measured by the cost to the NHS of that treatment, for example the patient has avoided the need for a treatment which may have been uncomfortable and permanently affected the integrity of a tooth.

Sensitivity analysis using fee-for-item costs

It could be argued that the English UDA system (with just three ‘bands’ of costs) could obscure cost differences as a result of lower treatment needs. Investigating the effect of using the Scottish fee-for-item costs was hypothesised as a sensitivity analysis, to provide a more discriminatory approach.

The intention here is not to estimate ‘equivalent’ Scottish NHS treatment costs and the values attached to each treatment item are not intended to accurately quantify resource use in absolute terms. It is included for comparison purposes of relative differences only.

The Scottish treatment item in the Scottish Dental Remuneration schedule 142 November 2019153 thought to best reflect the ‘middle’ level of complexity for each type of treatment recorded in the English FP17 clinical data set was chosen as the reference cost (Table 29). The reference costs are multiplied by the number of treatments for each patient recorded in the NHS BSA dental data set, to calculate the totals shown in Table 30.

Using a GLM and the regression results shown in Table 31, the difference in treatment cost is estimated as £9.89 lower per person in the optimally fluoridated group (95% CI −£9.39 to −£10.41). This is a relative reduction in treatment costs of 3.38% (95% CI 3.20% to 3.55%). Despite using a completely different philosophy to costing dental treatment, this estimate is remarkably similar to the relative reduction of 5.46% per person (95% CI −5.26% to −5.66%) we found in the main analysis (using average UDA values for England).

This provides reassurance that we are unlikely to have missed a large cost differential as a result of the banded approach to costing used in England, as compared to the more detailed fee-for-item costings used in Scotland.

Sensitivity analysis on units of dental activity values

In the main health economic analysis, the mean UDA value for contracts in England, obtained from NHS BSA, is used to calculate treatment costs. In Table 32 we present descriptive statistics using the UDA target, contract value and resulting UDA value calculated using contract specific values from the study sample. In each case, the mean is first calculated for each patient over all their courses of treatment, then the mean for both groups is calculated over all patients in that group.

Patients in the sub/non-optimally fluoridated group are more likely to have attended practices whose contracts have a very high financial value and a low UDA target. The resulting ‘UDA values’ (total financial contract value divided by UDA target) can be very high. It is likely that such contracts have a proportion of their financial value/activity allocated to a type of care that is not commissioned using UDAs. Orthodontic care is commissioned using UOAs, and sedation services are commissioned using specific KPIs, for example, number of patients treated, or waiting times. Community Dental Services are sometimes commissioned and funded to provide treatment under general anaesthesia in secondary care using a PDS (primary care) contract, with a very high contract value but very few UDAs allocated. We did not have access to these alternative commissioning/monitoring metrics, or the proportion of financial contract value that they apply to; therefore, it is not possible to know how much of the total financial contract value is actually allocated to UDAs, and what proportion is allocated to non-UDA-based activities.

Extreme differences in the contract-specific ‘UDA values’ as shown in Table 32 likely result from the fact that sub/non-optimally fluoridated patients are sampled in larger numbers and from a wider geographical catchment, to facilitate the one-to-three matching approach. Using the contract-specific UDA values contained within the study sample results in very extreme cost differences which do not appear to be driven by similar differences in clinical activity, but rather the different contracts and mechanisms available to commission dental services. This is why we elected to use an average UDA value obtained from NHS BSA to apply costs to clinical activity.

Objectives

1. To gain insights about what matters to patients and the public about their oral health when developing the original proposal for the bid.

2. To seek ongoing input and invite challenge into the conduct and governance of the study.

3. To gather views from patients and the public on what they would consider to be a meaningful difference in our primary outcome.

4. To develop dissemination materials that are engaging and clear.

Objective 1

In designing the research proposal, the views of 16 members of the public were sought using one-to-one conversations at a dental practice in a deprived location in central Salford. Volunteers were reimbursed for their time with a £10 gift voucher, funded by a bursary from the North West Research Design Service. This group were considered to be appropriate because they were drawn from the population of patients we were planning to gather data on, and they include a wide cross section of society with the potential to benefit from water fluoridation. With many having seen the dentist that day, they were also expected to be focused on the topic for the study application.

Objective 2

Embedded PPI members were recruited to both the OMG and the Study Steering Committee (SSC). The OMG met once every 2 months and the SSC met four times. The embedded PPI members were recruited through the NIHR ‘People in Research’ website. We budgeted for PPI involvement throughout the study, using the NIHR INVOLVE calculator. One staff member was assigned as the main contact person for both the PPI members of the team. This staff member was responsible for recruiting the PPI members, undertaking an initial training needs analysis, offering training options, introducing the PPI members to the study team, guiding, supporting and working with them throughout the study.

Objective 3

Three online workshops were held with a view to gathering views on what would be considered to be a meaningful reduction in the number of invasive NHS dental treatments received over 10 years, as a result of water fluoridation. Ten patient and public contributors were recruited through the ‘People in Research’ website, and eight were recruited from the existing University of Manchester Primary Care PPI group (PRIMER). The first 10 contributors to respond to the invitation to take part on the ‘People in Research’ website were included. The findings from the day were distributed to the interested participants via e-mail after the event. Further information on the content of and findings from the workshop are available in Chapter 3, Stakeholder views on a minimum ‘clinically’ relevant effect size.

Objective 4

All public-facing materials were codeveloped with the PPI member of the study OMG, and this chapter and the plain language summary of this report were written together. The full study report has also been reviewed by the PPI member. PPI members will be involved in developing further dissemination materials, and in sharing outputs through their wider networks and during our planned engagement workshop for policy-makers. We also plan to share the public-facing dissemination materials yet to be developed (the video animation and infographic) with groups from a range of backgrounds for comment and feedback during their development process.

Objective 1

These conversations revealed that researching ways to improve dental health was seen as very important because it is something that affects everyone. The out-of-pocket costs to patients of dental treatment were seen as a major issue, as well as the wish to avoid anxiety-provoking treatments. As a result, more information on the impact of patient charges and dental anxiety was added to the ‘background’ section of the funding application. The majority of people wrongly assumed they already had fluoride in their drinking water and seemed shocked to find that while some areas were potentially benefiting from this intervention, their area wasn’t. Some stated if this study identified a reduction in the number of unpleasant dental treatments required, or costs for patients, they would be asking their local council why they don’t fund water fluoridation.

The patients felt that reducing the need for dental treatment was important, firstly because it is unpleasant and, secondly, because it is costly to patients because of patient charges. These views helped to solidify our plan to use ‘invasive dental treatments received’ as the primary outcome measure, rather than the usual epidemiological index for dental studies, ‘number of decayed, missing and filled teeth’ (DMFT). One person suggested that a good way to share the results with the public would be to create a video which could be shared through social media. This was subsequently added to the dissemination package arising from the public feedback.

Objective 2

We had a very good response to the advertisement for embedded PPI contributors on the ‘People in Research’ website. The PPI member of the OMG had experience on committees and belongs to a number of networks, but aside from being a dental patient herself she had little existing knowledge about dental health. When applying for the role she was interested in the potential of dental research to help give a person a better QoL, free from dental pain, more confidence and a more positive self-image. She also hoped to increase her understanding about the potential health impact of water fluoridation. We developed a bespoke package of training, made up of a slide deck and key documents related to oral health. These helped to increase her confidence in contributing to this topic area and helped her to achieve her aim of increasing her understanding about the potential health impact of water fluoridation. This PPI member prepared a plain English summary for us to add to the current version of the protocol and helped to refine the text for the study website. She also assisted with developing the slides for the meaningful difference workshops. Through her involvement with this project, she has become more knowledgeable about dental issues and has also contributed to other funding applications within the dental school.

A second PPI member was recruited to the SSC. His training needs analysis revealed that he was a very experienced PPI contributor and was also a member of the University of Manchester Primary Care PPI group (PRIMER). He did not feel any particular training would be needed. His attendance at the steering committee meetings resulted in changes to the steering committee terms of reference and the study protocol.

Objective 3

A total of 18 public contributors attended the three workshops. The findings showed for most patients and the public, a minimum of one invasive dental treatment per person, over 10 years, would need to be avoided as a result of receiving water fluoridation for the effect to be viewed as meaningful (a 16% relative reduction). Interestingly, in contrast, dentists felt even very small reductions of less than half a treatment per person over 10 years (5% relative reductions) were meaningful. This finding was used to contextualise the results of this study and contributors reported they valued receiving a summary of the findings from the day and would like to continue to receive information about the study and our results. A negative incident in one of the workshops was that one contributor (while not ‘muted’) on the online call made a disparaging remark about another contributor. This was very unfortunate, and as it was not possible to identify which person had made the remark, it was not possible to remove them.

Objective 4

The public-facing study website, the plain English summary of the study protocol, this chapter and the full report have been codeveloped with or reviewed by the PPI member of the OMG. This has increased the readability of the documents and has also had additional benefits such as allowing us to see the particular elements of the research which are unique, surprising and most interesting to members of the public. Through the PPI member, the staff members of the research team have also learnt about patient engagement networks we were not aware of previously, that could be further routes for dissemination, such as the ‘use MY data’ and ‘Understanding Patient Data’ networks.

Primary objective

The findings of this study indicate the rate of invasive dental treatments in the optimally fluoridated group was 3% lower than that of the sub/non-optimally fluoridated group (IRR 0.969, 95% CI 0.967 to 0.971).

The predicted count of the number of invasive treatments from this model was 5.616 (SE 0.004) for the sub/non-optimally fluoridated group and 5.443 (SE 0.005) for the optimally fluoridated group. There was a predicted mean difference of 0.173 fewer invasive dental treatments per person (95% CI 0.185 fewer to 0.161 fewer) in the optimally fluoridated group compared to the sub/non-optimally fluoridated group.

Most stakeholders we engaged with (50/54) did not consider this magnitude of reduction in the number of invasive dental treatments received per person to be meaningful.

Secondary objectives

Strengths

There are several strengths to this research study. We published an a priori and peer-reviewed protocol and the study was designed to address some of the evidence gaps and key recommendations for water fluoridation research made by the MRC in 2002. 163 In particular:

• the effect of water fluoridation on the dental health of adults

• the effects of fluoridation on health and well-being beyond the usual measures of decayed, missing or filled teeth

• the economic impacts of water fluoridation

• how the effects of water fluoridation vary with social class.

There have been few studies on the effects of water fluoridation on the dental health of adults, because of the logistical difficulties in recruiting large number of representative adults into an epidemiological study. Routinely collected NHS dental claims data offered an innovative solution but have been a relatively underused resource for research, particularly in comparison to other NHS activity data sets, for example, Hospital Episode Statistics (HES), or the general medical practice-based Clinical Practice Research Datalink database. 165 In working with NHS BSA we have increased our knowledge and understanding of the potential for these data sets to be used for research. These reflections inform our recommendations for the further development of these data sets, and the utility of NHS dental data sets across the four UK home countries has been explored in more detail in a separate paper. 165 The benefits of using these routinely collected data are that very large samples are available and the data provide an ability to follow patients over long time periods. Once the necessary permissions are secured, the data are available at minimal costs and within a relatively short time frame. Important demographic variables are routinely collected, including age, sex, ethnicity and postcode. 165 Postcode is crucial as it enables the LSOA in which each patient lives and its level of deprivation to be assigned via the IMD index. This feature, and the large sample size, allowed us to address the MRC recommendation to investigate how the effect of water fluoridation deprivation varied with social class (patient deprivation decile and NHS exemption status used as proxy measures for social class).

Having access to the LSOA of each recorded patient address over the period of observation also allowed us to assign an individual level of exposure to water fluoridation when combined with address history recorded in the NHS BSA database. This element strengthens the natural experiment design of the study, which retains the unplanned, and essentially random variation in receipt of the intervention at the individual level. This feature helps to reduce the effect of area-based confounding related to factors which influenced decisions on which areas to fluoridate in the 1960s and 70s. While not perfect, the method of exposure assignment is much more robust than has been utilised in many previous studies on the effects of water fluoridation. We were also able to create an unplanned additional resource for future research, through the publication of LSOA level water fluoride concentrations. 128 The geodemographic variables in the NHS dental services data set also enabled sophisticated matching of exposed and non-exposed (comparator) samples. The propensity score matching procedure was carried out according to a pre-planned set of covariates which had been identified through the use of a DAG, in order to make the theoretical assumptions transparent and avoid the introduction of collider bias. An additional benefit of propensity score matching is that the method of addressing confounding is carried out before the final effect estimation (unlike with regression-based methods).

Patient, public and stakeholder involvement and engagement has been an important part of this project throughout and is a significant strength of this work. With regard to the MRC recommendation to consider the effects of water fluoridation on health and well-being beyond DMFT, we engaged with patients and the public to identify what was important to them about their oral health. They stated that avoiding invasive dental treatments (the ‘drill’ and the ‘injection’) was very important, as was keeping their natural teeth, and paying less in NHS dental charges. These conversations were not intended to provide a representative or generalisable view as they were involvement and engagement activities, but they did help to inform our decision to use number of invasive dental treatments as our primary outcome. We also hypothesised that this choice of primary outcome would have greater utility to be able to capture the effects of repeated treatments for decay on the same tooth (the restorative cycle), than the traditional measure of DMFT. The three online stakeholder involvement and engagement workshops we held also helped us to understand the range of effect sizes that patients, the public and dental experts would consider to represent a meaningful reduction in invasive dental treatments as a result of water fluoridation. The range of views we encountered was broad, and this was reflected even within the research team. It is possible that the views of readers and policy-makers may differ from the majority view that we found in our workshops.

Using a more formal approach, a minimum set of patient-centred outcomes for adult oral health has recently been developed in a collaboration between the World Dental Federation (FDI) and the International Consortium for Health Outcomes Measurement (ICHOM). 166 The authors report that the aim is for the outcome set to be used to measure the value of healthcare outcomes, in clinical, research, evaluation and community settings. 166 Although ICHOM is designed to capture the outcomes of healthcare services, rather than preventive interventions, the outcomes reflect what patients and the public said was important to them during those early conversation in Salford, including the ability to eat and smile. Interestingly, the initial iteration of potential outcomes for ICHOM did not include the financial impact of dental care on individuals, but this outcome was added as a result of feedback from an advisory group of eight patients, and aligns with what we heard during our engagement work. 166 Avoidance of irreversible treatment is an obvious goal for preventive interventions, but this was not included due to the nature of the ICHOM project, which is specifically to develop measures which represent end results of health care. 166 The core outcome measures in effectiveness trials (COMET) initiative collates core outcome sets specifically for intervention research purposes. To date, there remains no formally agreed core outcome set for the primary prevention of dental caries on the COMET database. 167 Developing one using the COMET methodology, which includes a systematic review of previous outcomes and Delphi consensus methods, is a necessary area of future research for caries prevention studies. 168

The health economic analysis provides contemporary information on the economic impacts of water fluoridation, from both the perspective of the public sector and NHS dental patients. We were able to source capital and revenue costs of water fluoridation centrally from PHE based on their responsibility for funding capital investment and recovering operating costs from local authorities, meaning they were accurate as they are what has been charged to the public sector. We also have high confidence in the costs to the NHS of dental treatment, as our data are based on claims that have already been processed and paid and are subject to independent scrutiny with regard to the detection of fraud. A sensitivity analysis based on the Scottish fee-per-item system of costings was also undertaken to provide a more discriminatory approach to costs than the simplified ‘banded’ cost levels used in the English system. Reassuringly, the relative differences in costs using these very different philosophies and approaches to costings were remarkably similar. We found a reduction in costs to the NHS of 5.46% per person (95% CI −5.66% to −5.26%) in the main analysis (using average UDA values for England), compared to a reduction of 3.38% (95% CI -3.55% to -3.20%) in the fee-per-item sensitivity analysis. Finally, we benefitted from independent oversight and scrutiny through the study steering group and were able to utilise their judgement and expertise when considering the most likely ROI and cost-effectiveness scenarios.

Limitations

There are several limitations to this work. It must be remembered that the study sample is drawn from the population of users of NHS dentistry living in England between November 2010 and October 2020. Population groups not included are those who prefer private dentistry, cannot get access to NHS dentistry, or have not felt the need for any dental care within this time frame. For NHS BSA to be able to link treatment claims together for a single, unique patient, a 1 : 1 match between NHS BSA Index ID and NHS number was required. The majority of treatment claims (FP17s) do not include NHS number when they are submitted, but NHS BSA are able to trace NHS numbers for 70% of submitted claims through the NHS demographic batching service. Published work using NHS HES data has demonstrated that missing data (and missed matches in NHS data linkage studies) more frequently affect ethnic minority patients, those living in deprived areas, foreign nationals and those with ‘no fixed abode’. 169,170 Missing data from specific population subgroups have the potential to bias estimates of effect, for example by eliminating higher need groups which could underestimate the effectiveness an intervention. More complete data on NHS numbers and ethnicity in the NHS BSA system would be welcomed to increase the ability to undertake data linkage. It must also be remembered that children aged below 12 years were not included in our cost-effectiveness estimates here and we were also unable to include societal benefits such as reduced time off work/school with toothache, attending dentist for treatment, reduction in pain and suffering from toothache, etc. Although these impacts fall outside of our public sector perspective, they do generate value for society.

A further limitation in the analysis of cost effectiveness is, as we were not able to secure an indicator of which areas were fluoridated as part of a public health programme, and which received natural fluoridation, we included all individuals who had a personal water fluoridation exposure > 0.7 mg F/l in the exposed group, regardless of source. While the proportion of the population receiving fluoridation through natural sources is relatively small (~5%),21 this means any benefits of natural fluoridation have been included in our analysis despite there being no costs associated with this exposure. This will have the tendency to overestimate the cost effectiveness of water fluoridation as a public health intervention. Additionally, we were not able to source capital costs earlier than 2013, which means the costs of setting up a new water fluoridation treatment programme have not been included in our cost-effectiveness estimates, as none were commissioned within that time period. These costs are potentially significant and include the costs of public consultations, feasibility studies, detailed engineering proposals and costs of new buildings, dosing equipment and service roads. Therefore, our approach considers the current impact of existing schemes. Individual local authorities can use our findings together with the setup costs they are presented with for any potential future water fluoridation investments. We provide information on the likely impact should a scheme be introduced and not the cost of its introduction. This is a pragmatic approach because the costs of investment will vary widely between settings because of the different engineering configurations required (water flows, road infrastructure and the arrangement of water treatment facilities). 149

The primary outcome measure of ‘invasive dental treatments received’ was chosen because it was stressed as being important to patients and the public during our engagement work. However, it has some limitations which are important to consider. Firstly, there is no financial incentive for dentists to accurately record the number of treatment items provided on the FP17 form (beyond one treatment in each ‘band’ of care). On the other hand, the number and types of treatment provided are automatically recorded from the electronic dental clinical notes system. As the regulator of dentists in the UK, the General Dental Council states that dentists must ‘make and keep contemporaneous, complete and accurate patient records’ to comply with the expected standards of the profession. While there may be some inaccuracies in the reporting of the number and types of treatments, we do not feel that this is likely to be a significant problem. Secondly, receipt of dental treatment is influenced by a range of factors unrelated to water fluoridation and dental caries, including the clinical decision-making of dentists, the presence of non-caries-related dental problems such as periodontal disease and trauma, and the patient’s ability to pay privately for some treatments (dental care provided privately following an NHS examination will not be captured). In the same vein, for the DMFT and ‘missing’ teeth outcome, we do not have information on the reasons why teeth were extracted, and some will have been extracted as a result of periodontal disease rather than caries. We have controlled for sociodemographic factors that may be related to non-caries-related treatments including periodontal disease and ability to pay privately (matching on age, sex, deprivation, ethnicity and NHS dental charge exemption status); therefore, we see no reason why non-caries-related treatments should vary between groups. Any relative differences in caries-related treatments would still be expected to have been apparent ‘over and above’ the non-caries-related treatments, but it is important to remember this measure is not a measure of ‘caries-related treatment’. The inclusion of a diagnosis code within the routinely collected NHS dental data sets would be a welcome step to provide greater sensitivity in future research studies and would aid epidemiological surveillance.

The DMFT outcome did not demonstrate a clear social inequalities gradient, which inevitably raises questions regarding the validity of these data. However, recent secondary analyses of UK and Australian epidemiological surveys have indicated that measures of total caries experience (DMFT/S) may not be sensitive to socioeconomic inequalities in adults, other than in the youngest adult age groups, and that missing teeth are more sensitive. 6,66,73 Indeed, the missing teeth outcome did demonstrate a clear social inequalities gradient in the present study. It should be remembered that DMFT in this study is not the same measure as the commonly used epidemiological index. In the NHS BSA data set, the number of DMFT variables are automatically extracted from the dental charting completed at the time of examination. 171,172 NHS BSA guidance for dentists states that teeth should only be recorded as missing if the dentist is ‘confident that the tooth was extracted’. This is in contrast to recent UK epidemiological surveys of adult oral health where any tooth not present in the mouth is charted as missing (i.e. including unerupted or congenitally missing teeth). 171–173 The accuracy of the DMFT data in this clinical context depends on the thoroughness of the dentist’s examination and the accuracy of their record keeping (as opposed to training and calibration as is used in epidemiological surveys). Previous audits have shown that the accuracy of NHS dental charts is variable when compared with a reference standard. 174

It cannot be ruled out that the absence of a clear effect of water fluoridation on DMFT or missing teeth may in part be due to the potentially lower sensitivity of clinical chartings carried out by busy GDPs, as opposed to trained and calibrated examiners. On the other hand, dental practitioners are able to dry the teeth, take radiographs and have access to more powerful lighting than is used in epidemiological studies, where examinations are often carried out in schools; thus, it could be argued that the sensitivity of the clinical examinations may be higher. Missing teeth are also a fairly obvious feature which do not require a thorough dental examination to detect. It seems unlikely that if the effect of water fluoridation had remained the same as previously reported (26% relative reduction in DMFT in the Cochrane systematic review),47 that it would go undetected solely as a result of a lack of sensitivity in these measures. It is also reassuring that the proportion of adults estimated to be edentulous in our full unmatched cohort (1.3%) is very similar to the proportion recorded in the 2017 National Dental Epidemiology Programme survey of adults attending general dental practices which used calibrated examiners (1.1%). 175 However, a study to investigate the validity of the DMFT data recorded by GDPs and transferred to NHS BSA as compared to trained and calibrated examiners would help to unpick this further.

For the assignment of individual water fluoridation exposure, we linked water fluoride concentrations between 2010 and 2020 to each patient’s residential history as recorded at the time of each course of treatment. Errors of misclassification could have occurred during the mapping of water fluoridation, due to the conversion from WSZs to LSOAs, and as a result of patients moving address in between courses of treatments. NHS BSA dental data are retained for a maximum period of 10 years; therefore, we were not able to assign water fluoridation exposure to individuals earlier than this date. Thus, the effects observed cannot be said to result from only 10 years of exposure, rather a ‘minimum of’ 10 years exposure. Additionally, measurement of exposure was concurrent with measurement of outcomes in this study, but reverse causality is not a concern as we see no mechanism whereby the number of treatments received could influence a patient’s fluoridation status. The variability in achieved water fluoride concentrations over time is such that it is important to have this information when assigning exposure to water fluoridation. However, the method we needed to use to obtain this information [sourcing water quality sampling data from each water company and converting the information to useable (LSOA) geographies] was very time-consuming and not one we would recommend to other research teams.

As with any observational study, the likely effect of unmeasured and residual confounding must be considered. Unmeasured confounding results from confounding that cannot be taken into account during the design or analysis stage of the study as the variables cannot be, or were not, measured. We did not have any data on the behaviours of participants, for example, tap water consumption, smoking, sugar consumption, fluoride use and oral hygiene. However, as exposure to optimally fluoridated water does not rely on individuals seeking out a particular service or intervention (as with many other health promotion interventions), the potential for confounding due to behavioural factors is significantly reduced. Behavioural characteristics of individuals are unlikely to have also influenced their group assignment. We propose that patient behaviours could only plausibly be linked to water fluoridation group assignment through the effect of sociodemographic characteristics clustering by geography and influencing decisions on which areas to fluoridate in the 1960s (which we accounted for by matching on age, rurality, ethnicity, deprivation and NHS charge exemption status). We also propose that data on sociodemographic characteristics are more likely to be accurate than self-reported behaviours.

An example of a potential source of unmeasured confounding is if there were historic ‘cultural’ variations in dietary or oral hygiene behaviours (not explained by age, rurality, ethnicity, deprivation and NHS charge exemption status) that influenced decisions on which areas to fluoridate in the 1960s, and which still persist today and influence oral health outcomes. We would not be able to remove the confounding effect of ‘culture’ because it is unmeasured. Concerns about unmeasured confounding are reduced in this study because of the natural experiment study design, which (through individual-level assignment of receipt of optimal water fluoridation) reduces the correlation between historic decisions to implement water fluoridation and group assignment (due to breaks in water fluoridation supply, suboptimal dosing, or individuals moving into and out of fluoridated areas). While still an observational study, there is some unplanned variation in receipt of the intervention, which takes the group assignment of individuals in this study closer to being ‘as-if randomised’, strengthening causal inference and reducing the likely impact of unmeasured confounding. 101,102

Residual confounding is when the variables included do not fully remove the confounding effect because they are poor proxies or are poorly measured or completed. The principal areas of concern in this study relate to socioeconomic status and ethnicity. Our analysis used two methods of accounting for socioeconomic status: an area-based deprivation measure (IMD) linked to the individual patient via postcode and NHS exemption status at baseline. It is possible that some residual confounding could occur as a result of facets of socioeconomic status not captured by these measures, for example through level of education, or as a result of people of higher socioeconomic status living in deprived areas (or vice versa). This could limit our ability to fully account for the effect of socioeconomic status. The categories used for ethnicity in this study are also relatively broad, which could limit the ability of this measure to truly account for the impact of ethnicity for all ethnic groups. Furthermore, we did not have any ethnicity information recorded during any course of treatment for 20% of our sample. While the direction and effect of ethnicity on oral health outcomes in the UK are not clear at present, it is likely that these missing data could have led to residual confounding due to ethnicity. 176 Greater completion of ethnicity data in NHS dental records is needed.

We did not take account of variation in UDA values across different contracts, and how this may influence the behaviour of dentists. This was because we did not have access to any other KPIs that those contracts were commissioned to deliver; therefore, UDA values calculated by dividing total financial contract value by UDA target cannot be assumed to be reflective of the true payment: targets ratio of any given service. Furthermore, operating costs for dental surgeries do vary in different regions of the country; therefore, it is not possible to know how much of the observed variation in UDA values is actually appropriate given the variation in overheads.

Introduction

This study is not a traditional evaluation of water fluoridation. Historically, epidemiological evaluations of dental caries prevention interventions usually use clinical measures of disease (DMFT/S) as the primary outcome measure and try to recruit representative samples of the exposed and unexposed comparator populations. Follow-up of the exposed and comparator samples is usually prospective and ideally for a minimum of 3 years. 47 The MRC (2002) report on water fluoridiation163 and systematic reviews of water fluoridation point to the lack of high-quality studies examining the effect of water fluoridation on caries in adults. 47 The rationale for our approach in using NHS activity data is that it enabled an affordable, contemporary and initial assessment of the effect of water fluoridation on an adult population in England.

While the clinical results suggest very small effects of water fluoridation that stakeholders may not consider to be meaningful, the health economic analysis suggests that water fluoridation represents a positive ROI for the public sector. These apparently conflicting findings can be understood in the following ways. Firstly, the ROI analysis used total NHS treatment costs and includes all types of NHS dental treatment, not just invasive dental treatments (e.g. urgent dental care). Secondly, the per-person reduction in overall NHS dental treatment costs was of a similarly small magnitude to the clinical results in relative terms (5.5% lower NHS treatment costs compared to a 3% lower rate of NHS invasive dental treatments). However, because the costs of NHS dentistry are so much higher than the costs of water fluoridation (£402.22 vs. £10.30 per person over 10 years in 2020 prices), even relatively small reductions in NHS dental service utilisation will result in positive returns. Finally, it must be remembered the estimated ROI is for the period of observation only and does not include the significant capital investment that would be required if a new water fluoridation programme was established. It was not possible to include the original set-up costs for the schemes currently operating in England because they were commissioned around 60 years ago.

If water fluoridation continues to provide effective disease suppression in adults at a similar level to that reported for children in the most recent Cochrane systematic reviews (mean DMFT 1.16 teeth lower per person in the optimally fluoridated group (95% CI −1.61 to −0.72)),47 one would expect that a large difference in invasive treatments between exposed and non-exposed populations would be readily found. The observed reduction in the number of invasive dental treatments received per person by those in the optimally fluoridated group was smaller than what the majority of stakeholders we consulted with (patients, the public and dental professionals) felt would be meaningful. The stakeholders were drawn from both within the North West and across the UK, but it is important to remember that the workshops were an involvement and engagement activity rather than an attempt to produce a generalisable cut-off. Nonetheless, relative effect sizes of 3% or less have previously been referred to as ‘exceedingly small’ effects, which warrant cautious interpretation as they ‘could be eliminated with even minimal biases and their importance is uncertain’. 161

Potential explanations for the failure to find a definitively meaningful difference in clinical outcomes between exposed and non-exposed groups include:

1. a smaller effect size for water fluoridation in contemporary populations in which fluoride toothpaste use is ubiquitous than has been previously reported in historical studies (without widespread use of fluoride toothpaste and higher disease levels) and by systematic reviews made up predominantly of historical studies. This is supported by more recently published studies as outlined in How the results sit with the wider literature.

2. lack of a clear relationship between individual levels of dental disease and treatment provision due to perverse contracting incentives within NHS dentistry which provide no additional remuneration over and above one treatment within each payment ‘band’.

While we feel it is unlikely that we have missed a large/meaningful effect, we cannot rule out some underestimation of the effect of water fluoridation, as a result of using data from NHS dental services. The current contracting model for dental services incentivises dentists to modify their treatment patterns to hit contract activity targets (which are not aligned to population need). 94,178 Dentists working in deprived areas with patients that have a higher need for care are required to undertake more clinical work for the same UDA credit. Under this contracting system, there is no financial incentive for a dentist to provide three fillings compared to one. This disincentive to ‘find’ both patients and teeth that need treatment could have resulted in an underestimation of the true effect of water fluoridation. There are major concerns about NHS dental services and it is now over 13 years since publication of the Steele Review and the promise of a new NHS dental contract in England. 179 A fresh consideration of how the resources available for oral health should be allocated to provide a service based on population need is long overdue.

1. Limitations in our ability to detect a meaningful difference as a result of using routinely collected and uncalibrated data for outcomes.

We do not feel that this is the case, because we were able to detect a meaningful difference in the number of invasive NHS dental treatments received as a result of deprivation (1.285 more per person comparing the most to the least deprived decile). Similarly, we were able to detect a difference in the number of missing teeth per person as a result of deprivation (1.19 more per person comparing the most to the least deprived decile). This indicates that with this very large sample size, these outcome measures are sufficiently sensitive to detect even small effects (i.e. the finely graded small differences between each decile), when they were present.

With regard to generalisability, there are some slight differences in the characteristics of our analytical (matched) cohort compared to the general population of NHS dental service users as a result of the matching process (and the small proportion of the English population receiving water fluoridation). In the interests of increased internal validity, the matched analytical cohort was made more similar to the optimally fluoridated population of England than the sub/non-optimally fluoridated population. Our matched analytical cohort is therefore slightly more urban, more ethnically diverse, more deprived and has a lower proportion of NHS fee-paying patients than the sub/non-optimally fluoridated population of England. 180 Despite these demographic differences, the clinical characteristics of the unmatched and matched cohorts were very similar and we are confident that our findings are generalisable to the wider population of NHS dental service users in England (the 2009 Adult Dental Health Survey found that 58% of adults had tried to make an NHS dental appointment in the previous 3 years and of these 92% successfully received and attended an appointment). 181

Implications for policy, public health practice and research

This is one of the few studies on water fluoridation that has included adults and adolescents. Using routinely collected data and an innovative design meant it was possible to address some of the MRC research priorities for water fluoridation which were first articulated in 2002. However, triangulation and the use of ‘interlocking evidence’ are recommended in guidance on the evaluation of population health interventions. 101,182 The results of this study and the recently published NIHR CATFISH study in children both support the hypothesis that water fluoridation appears to be producing less dramatic improvements in oral health in contemporary UK populations than it did in historical studies. Fluoride toothpastes and professionally applied fluorides are now ubiquitous, and as a result the prevalence and severity of dental caries have greatly reduced compared to when the majority of studies were carried out. Similar findings of small clinical differences in absolute terms have been found in recent studies of adult populations in Australia, the USA and Sweden. 62,66,70 Water fluoridation may now represent a classic ‘prevention paradox’; a preventive measure which can bring benefits for populations and services but ‘offers little to each participating individual’. 183 This was not the case when water fluoridation was originally discovered and implemented, when the early studies identified noticeable reductions in caries of around 50% per person. 184,185 Whether the case for water fluoridation can be made based solely on the potential for reductions in NHS dental service utilisation and savings in NHS treatment costs, rather than health gains that are noticeable and meaningful to patients and the public, needs to be considered by stakeholders. Participatory consultation methods, for example, citizen juries, may be beneficial to explore this issue further.

Decision-making context must be considered when evaluating the transferability of our findings internationally. Factors of relevance include the prevalence and severity of dental caries in the population, access to alternative sources of fluoride and universal preventive dental care, costing arrangements for both dental care and water fluoridation, and current coverage of community drinking water and water fluoridation. Settings with high caries rates and limited access to preventive dentistry and alternative sources of fluoride may find that water fluoridation is still producing large health impacts in their population. Decisions regarding expansion of water fluoridation versus the cessation of existing schemes will also need to be approached very differently. Both the present study and the recent CATFISH study of the effectiveness of water fluoridation in children support the conclusion that existing schemes in the UK are cost-saving. 164 On the other hand, new schemes can require significant capital investment which may not be recovered during their lifespan in (increasingly) low caries populations.

The costs of any new water fluoridation programmes are extremely variable and depend on local water supply configurations. 186 Any cost estimates for new water fluoridation programmes or significant investment in existing schemes should be viewed in the context of our estimate that receipt of optimal water fluoridation (≥ 0.7 mg F/l) over 10 years reduced NHS dental treatment costs by 5.5% per person. Breaks in the supply of optimally fluoridated water and suboptimal dosing are relatively common103 and would result in a smaller impact of water fluoridation programmes on NHS treatment costs. To provide some context on the potential costs of a new scheme, a 2009 feasibility report estimated that the capital works to establish a programme to fluoridate the whole of the North West of England would be £102,124,677 in 2007 prices (± 30%). For Greater Manchester and Merseyside the estimate was £35,447,837 (± 30%). 149 In 2021 prices, these estimates would be £139,288,507.19 for the North West and £48,347,534.05 for Greater Manchester and Merseyside. 187 We received estimates on the projected lifespan of such equipment from two of the five water companies that supply fluoridated water in England. Based on their experience, the fluoridation dosing equipment was estimated to last between 15 and 20 years and the civil infrastructure (e.g. building, roads, pipes) was estimated to have a lifespan of 60 years.

There is no doubt that population-level, ‘mass preventive’ interventions for dental caries are still required. While disease levels have fallen dramatically in high-income countries over the last 40 years, dental caries remains almost universal by adulthood, even in populations that have had access to fluoride toothpastes and fluoridated water from birth. 5,38 We may be reaching the limit of improvements that can be achieved through fluorides alone. A dose–response relationship between free sugars and dental caries is evident at all levels of intake above zero and fluorides merely attenuate this relationship. 8,36–39 Average consumption of free sugars in the UK is more than double the recommended level for adolescents, and it is almost double for adults. 9,10 Only 7% of adolescents aged 11–18 and 17% of adults aged 19–64 meet the recommendation that free sugars should make up no more than 5% of total energy intake. 10 In populations with widespread access to fluoride toothpastes and topical fluorides, policy interventions related to the social and commercial determinants of health which change the food environment, improve diet quality and reduce sugar consumption may offer greater improvements for the dental health of populations and individuals than further expansion of water fluoridation. 188 Dental and dental public health professionals who seek to prevent caries have a great deal in common with those seeking to promote healthy weight and prevent other nutrition-related chronic diseases. 15 These relationships have been underdeveloped in the past but working together on sugar reduction may help us to move beyond the fluoride era and achieve further advances in caries prevention.

Recommendations for future research

1. National Health Service primary care dental data are a potentially valuable resource for research, surveillance and epidemiology. Consideration should be given as to how to develop this resource further. Suggestions include: commissioning enhanced data collection through a network of dental providers (similar to the medical G.P. Clinical Practice Research Datalink network), mandatory use of NHS numbers, recording the reason for treatment (caries, periodontal disease, trauma, repair of old restoration), steps to increase the completeness of patient ethnicity data and validation of the DMFT data.

2. Routinely collected water quality monitoring data should be made publicly available. A central, annually updated database of water fluoride concentrations (mg F/l) at small area level is required to support future research and monitoring of water fluoridation.

3. To broaden the evidence base for triangulation, the effect of water fluoridation on the dental health of adults should be investigated in any future national dental epidemiology surveys.

4. To determine if the estimated cost to avoid one invasive dental treatment (the ICER) represents good value for money, further work is required to identify a threshold level for ‘good value’ for this outcome.