Data sources

This study used the French National Health Data System (SNDS, Système National des Données de Santé), which contains, for over 99% of the French population, individual-level medico-administrative data on outpatient services reimbursed by the National Health Insurance as well as hospital discharge diagnoses and long-term diseases (LTD) registrations (e.g., for cancer) [11, 12]. Drug claim data include dispensing dates for reimbursed medications together with their French presentation identification codes. The SNDS also provides demographic information such as sex, dates of birth and death, and place of residence. Hospital discharge diagnoses include “main” (e.g. myocardial infarction, chemotherapy) and “related” (e.g. breast cancer) diagnoses, as well as “associated” diagnoses, i.e. patient’s comorbidities involving a higher burden of care. All data can be linked using a pseudonymized identifier. The SNDS integrates data from several insurance schemes, the “General scheme” for salaried workers covering 75% of the population with data recorded since 2006. Smaller schemes have been progressively incorporated.

We also used data from the French Theriaque database (https://www.theriaque.org) as of 2021, which provides, for all pharmaceutical products ever marketed in France, detailed information including presentation identification codes, the list of excipients—including TiO2—, pharmaceutical form, strength, Anatomical Therapeutic Chemical (ATC) codes, and package dispensing unit numbers. Data on exact TiO2 quantities, obtained from the French medicines safety agency (ANSM) as of 2021, were available for 42% of TiO2-containing pharmaceutical specialties [2].

Selection of drugs

To minimize confounding related to participants’ characteristics (see Discussion), we restricted our analyses to users of selected drugs. Drugs were characterized by their active molecule, strength, and pharmaceutical form, and were selected based on these criteria:

1.

Availability in both TiO2-containing and TiO2-free formulations.

2.

At least 300,000 boxes of both TiO2-containing and TiO2-free formulations reimbursed annually between 2006 and 2016. This threshold was determined a posteriori to ensure sufficient precision in estimating the TiO2-cancer associations (see Results) while limiting the number of selected drugs.

3.

No excipient was exclusively present in either TiO2-containing or TiO2-free formulations, allowing us to disentangle the effects of TiO2 from effects of other excipients.

Applying these criteria, we initially selected: metformin (ATC code A10BA02) tablets at doses of 500, 850, and 1000 mg, tablets combining levonorgestrel 150 µg and ethinylestradiol 30 µg (G03AA07), tablets of acebutolol (C07AB04) 200 mg, ibuprofen (M01AE01) 200 mg, and doxycycline (J01AA02) 100 mg (see also Supplementary Methods and Table S1 in Online Resource). We excluded ibuprofen and doxycycline due to their predominant short-term use, limiting relevance for dose-response analyses. Additionally, the levonorgestrel-ethinylestradiol combination was excluded because the presence of TiO2 and carnauba wax were strongly correlated, precluding isolation of TiO2’s effects.

We combined all doses of metformin tablets to create a unique cohort of metformin users. Ultimately, the selected drugs were:

1.

Metformin tablets (all doses).

2.

Acebutolol 200 mg tablets.

Cohorts of selected drug users

For each selected drug (metformin and 200 mg acebutolol), we assembled a cohort of users by including individuals affiliated with the “General scheme” with at least one claim for the selected drug in the SNDS over 2006–2021. Follow-up for cancer incidence began on January 1, 2013, or the date of the first identified delivery of the selected drug, whichever was later. The earliest follow-up start date was January 1, 2013, because individuals in the “General scheme” who died before 2013 may not have had their unique identifier integrated into the SNDS [11]. Follow-up ended at the earliest of death, first recorded cancer diagnosis, or December 31, 2021.

We excluded individuals with a cancer diagnosis or death before follow-up started, and those whose first claim for the selected drug occurred after follow-up ended. The acebutolol 200 mg and metformin cohorts comprised 587,164 (Figure S1 in Online Resource) and 3,155,100 (Figure S2 in Online Resource) individuals, respectively.

Identification of cancer cases

Cancer occurrence was identified by:

a hospital discharge diagnosis with an ICD-10 code for cancer (codes starting with ‘C’ or ‘D0’), considering “main”, “related”, or “associated” diagnoses, or

an LTD registration with a cancer ICD-10 code.

The retained date of cancer diagnosis was the earliest of either an LTD registration for cancer or a hospital discharge with a cancer code.

Cancer cases were classified by site using ICD-10 main categories. Subcategories were also used (e.g., ovarian cancer within the “Malignant neoplasms of female genital organs” main category), provided they accounted for at least 15% of cases within the main category (Table S2 in Online Resource).

Selection of cases and controls

We implemented a nested case-control design within each cohort of selected drug users.

For each cancer case diagnosed during follow-up, up to ten controls were randomly selected using incidence density sampling. Controls were required to be cancer-free on the case’s diagnosis date (index date) and to meet the following matching criteria: date of first delivery of the selected drug identified in the SNDS (± 6 months), year of birth, and sex. Cases could serve as controls prior to their diagnosis. Thereby, the odds ratio is an unbiased estimate of the incidence rate ratio (RR) that would be obtained in a cohort study conducted within the source population [13].

In analyses where TiO2 exposure was characterized by cumulative dose rather than the cumulative number of tablets, cases with an unknown cumulative dose at diagnosis (i.e., if a drug box with an unknown TiO2 amount had been purchased at any time before diagnosis) were excluded. The same matching procedure as above was applied, with the additional requirement that controls had a known cumulative dose at the case’s diagnosis date.

The corresponding flow charts are shown as Figures S1 and S2 (Online Resource).

Exposure to titanium dioxide

Deliveries of the selected drug (metformin or acebutolol 200 mg, depending on the cohort of origin of the individual) were identified in the SNDS using French presentation identification codes. Cumulative TiO2 exposure was calculated up to five years before the index date (5-year lag), to avoid considering exposures unlikely to influence cancer risk, given cancer development latency periods [14]. Individuals whose first delivery of the selected drug occurred less than 5 years before the index date were excluded. Cumulative exposure to TiO2 was quantified as:

The number of TiO2-containing tablets of the selected drug reimbursed from drug claims data availability (2006) to five years before the index date.

The cumulative TiO2 dose (in mg) in the selected drug tablets reimbursed during the same period.

Covariates

Adjusted analyses included the French geographical social deprivation index [15] and geographical region (Table 1, and Table S3 in Online resource) at the time of the first selected drug claim. We also calculated the cumulative number of selected drug tablets containing the individual constituents listed in Table 1, from 2006 to five years before the index date.

Table 1 Main characteristics of cancer cases and matched controls