6. Analysis

6.1 Definition of the primary endpoint

The primary endpoint is breast cancer mortality, based on deaths coded by the Office of National Statistics (ONS) as having breast cancer (ICD-10 C50 or ICD-11 2C6) as the underlying cause. Thus, a death will be considered to have been from breast cancer if and only if breast cancer had been coded as the underlying cause in the final ONS death registration records. The primary analyses focus on breast cancer deaths in particular time periods (Section 6.3.3).

6.2 Definition of the main subsidiary endpoints

The main subsidiary endpoints are:

• invasive breast cancer ( ICD-10 C50 or ICD-11 2C6), from NHS England cancer registration records (overall, and subdivided by tumour characteristics such as diameter, grade, ER status and stage);
• in situ breast cancer (ICD-10 D05 or ICD-11 2E65), which may result in surgery or radiotherapy, again from these cancer registration records;
• use of mastectomy (OPCS-4 B27) that was recorded in the NHS England hospital episode statistics (HES);
• use of systemic breast cancer therapy, particularly breast cancer chemotherapy (OPCS-4 X70-3, X352, X373, X384, Z511-2), that was recorded in these hospital episode statistics or recorded in the NHS England systemic anti-cancer therapy (SACT) dataset; and
• use of radiotherapy for breast disease (ICD10 C50, D05, D24, N60-64 or ICD11 2C6, 2E65, 2F30, GB20-23) that was recorded in the NHS England radiotherapy data set (RTDS).

If NHS general practice records become conveniently linkable then any relevant information from them could eventually also be incorporated into the AgeX analyses of breast cancer treatments given, but treatment outside the NHS may well never be captured.

6.3 Methods for the primary analyses

After the exclusions described in Sections 4.2 and 5.2, the primary and main subsidiary analyses will be of the effects of actually having one additional screening visit. Parallel analyses will be provided of the effects of being randomly allocated to receive one additional screening invitation (Section 6.3.2).

6.3.1 Adherence-corrected analyses of the effects of one additional screening VISIT

Estimates of the effect of one additional screening visit on breast cancer mortality will use the method of Cuzick et al³ to allow for non-adherence, which avoids assuming similarity between those adherent and non-adherent to the random allocation. The aim of adherence-corrected analyses is to estimate the actual effects of a screening visit among those women who would have had a screening visit if, and only if, randomly allocated to be invited for screening. (Screening does not include mammography after a breast cancer has been diagnosed.)

Some women, who may well be atypical in unknown but relevant ways, would not get screened, regardless of what their random allocation happened to be. Some other women, perhaps atypical in other relevant ways, would get screened anyway, again regardless of what their random allocation would be. All others would have a screening visit if, and only if, randomly allocated to be invited. *

Adherence-corrected analyses of the effects of actually having, vs not having, a screening visit involve comparing (A – B) vs (D – C), with groups A, B, C and D as defined in the table. For, subtraction of group B is equivalent to removing from group A those who would have been screened even if not invited, leaving those who would have a screening visit if, and only if, invited. Likewise, subtraction of C is equivalent to removing from D those who would not have been screened even if allocated to be invited, again leaving those who would have a screening visit if, and only if, allocated to be invited. *

This comparison of (A – B) vs (D – C) assesses unbiasedly the full effects among those who would have a screening visit if, and only if, allocated to be invited, without assuming that groups A, B, C and D are comparable with each other.³ This method of adjustment for adherence has little or no effect on the statistical power to detect any differences in breast cancer mortality between invitees and controls.

_______________

* There might also, at least in theory, be a tiny group who, perversely, would all get exactly the opposite of whatever their random allocation happened to be. Ideally these non-adherent women would be excluded when analysing the results, but as they are not individually identifiable they cannot be. This does not, however, introduce any bias at all into the adherence-corrected estimates of the proportional effect of screening (as the adherence correction subtracts this tiny non-adherent group from both sides of the comparison).

6.3.2 Intent-to-treat analyses of the effects of one additional screening INVITATION

Parallel analyses will be provided of the effects of being randomly allocated to receive one additional screening invitation. These parallel analyses are modified intent-to-treat (mITT) analyses; “modified” indicates restriction to the 3 million included invitees and controls, rather than all 4.5 million women originally randomised.

6.3.3 Focus on appropriate time periods

To assess the eventual effects of one additional screening visit (or of one additional screening invitation) on breast cancer mortality, follow-up must be long (a decade or two by the time of the final analyses, depending on the year of recruitment) and the analyses of breast cancer mortality must be strictly unbiased and as sensitive as possible, focusing on the periods in which an appreciable effect can reasonably be expected.

No material effect of the random allocation should be expected on breast cancer mortality during the first few years after randomisation. Hence, the primary analyses are restricted to the later breast cancer deaths (Sections 6.3.4 to 6.3.5).

A few years after just one single additional screening invitation, the annual incidence rates of newly diagnosed breast cancer may well become similar in invitees and in controls. No material effect of the random allocation should be expected on mortality from breast cancers that are diagnosed after this convergence of incidence rates. Hence, the primary analyses are further restricted to deaths from breast cancers that had been diagnosed only a few years after randomisation (Sections 6.3.4 to 6.3.5).

6.3.4 Tabulation of breast cancer deaths in younger women

In the younger women (randomised at ages 47-49) the annual incidence rates of new breast cancer became similar by year 4 after randomisation, because by then all the invitees and all the controls should have been invited recently for routine screening (at ages about 50-52). This routine screening invitation makes the cumulative incidence in controls catch up with that in invitees within 4 years of randomisation.

The numbers of breast cancer deaths among younger women will therefore be tabulated, with and without adherence correction, both by years from randomisation until diagnosis of the fatal cancer and by age at death, in the format of Table 2. The adherence-corrected tabulation addresses the effect of actually having one additional screening visit, and the parallel (mITT) tabulation, without adherence correction, addresses the effect of having been allocated one additional screening invitation. Analyses of both of these tabulations will focus on mortality after reaching age 55 from a breast cancer known to have been diagnosed < 4 years after randomisation, with the primary analysis being that of the adherence-corrected table. Dates of diagnosis will be taken from the linked national datasets (primarily using cancer registry data, but using other datasets as registry data if these provide a date of diagnosis and the registries do not).

In Table 2, hypothetical numbers of breast cancer deaths (not adherence-corrected) are inserted into the subtotals for mortality from breast cancers diagnosed during the first 4 years after randomisation and those diagnosed later (or at an unknown time). These illustrate what the final results after follow-up to 12.2031 might be if random allocation to one additional screening invitation at ages 47-49 reduced by about 20% the probability of having a breast cancer diagnosed < 4 years after randomisation that caused death after reaching age 55 (shaded parts of Table 2) but had no effect on the other breast cancer deaths.

Regardless of whether such a 20% difference is considered plausible, any effect of one additional screening invitation on breast cancer mortality may well chiefly affect just the results in the shaded parts of Table 2. The same would be true of the adherence-corrected version of Table 2. Hence, the primary analysis will focus just on the sum of the adherence-corrected results in the shaded parts.

Three-yearly mammographic screening detects a higher proportion of all ER+ breast cancers than of all ER– breast cancers. Hence, subdivision of the breast cancer deaths by ER status (ER+, ER– or unknown) will also be tabulated in the same format as for all breast cancer deaths, but chief emphasis will be on the overall breast cancer mortality results, regardless of ER status.

6.3.5 Tabulation of breast cancer deaths in older women

In the older women (randomised at age 71-73), no further invitations were scheduled for the controls. Hence, the time when the annual incidence should be expected to become similar in invitees and controls cannot be predicted as easily as in the younger women. Nevertheless, lacking further information, the analyses of breast cancer mortality in the older women will involve the same methods and format as those in the younger women, except that the age groups will be 70-74 (when little effect on breast cancer mortality should be expected), 75-79, 80-84 and 85+.

6.3.6 Kaplan-Meier plots

For both younger and older women, adherence-corrected Kaplan-Meier plots of cumulative event rates by time since randomisation will be given for death from a breast cancer diagnosed < 4 years after randomisation, and for death from a breast cancer diagnosed later (or at an unknown time). These describe the effects of one additional screening visit. Parallel results will also be given for mITT analyses of the effects of one additional screening invitation (without correction for adherence).

6.3.7 Summary of the definitions of the primary analyses

The interim and final primary analyses will include only the 2 million younger and 1 million older participants who had been eligible and linkable to NHS England datasets, with no record there of prior cancer or breast disease, and who were considered by pre-defined criteria likely to attend screening if invited (Section 5.2; Table 1).

These will be adherence-corrected analyses of the effect of one additional screening visit (Section 6.3.1). Parallel “mITT” analyses will also be given of the effect of being allocated one additional screening invitation.

The primary analyses will be restricted to deaths from a breast cancer diagnosed < 4 years after randomisation that occurred after reaching age 55 for younger women (Section 6.3.4; Table 2) and after reaching age 75 for older women (Section 6.3.5).

6.4 Methods for the main subsidiary analyses

For both younger and older women, adherence-corrected Kaplan-Meier plots of cumulative (first) event rates by time since randomisation will provide the main subsidiary analyses of the effects of one additional screening visit on breast cancer incidence and on the eventual use of mastectomy, of radiotherapy, and of systemic therapy (particularly chemotherapy). Again, parallel results will also be given for mITT analyses of the effects of one additional screening invitation (without correction for adherence).  The final primary and subsidiary analyses will be accompanied by appropriate health economic analyses.

6.5 Missing data

As record linkage is reliable and the endpoints are defined to be of the recorded numbers of events in the linked national datasets there can be no missing data for the primary and main subsidiary endpoints, except for those women no longer available for record linkage (see Section 5.3).

6.6 Allowance for clustering in analyses of breast cancer mortality

Wherever numbers of deaths from breast cancer in invitees and in controls are to be compared, calculation of the statistical significance of the comparison will make appropriate allowance for the cluster randomisation. This will make little difference to calculations of statistical significance or of confidence intervals, as the average numbers of breast cancer deaths per cluster will be small even in the final analyses.

For the hypothetical final numbers in Table 2 for younger women the average number of breast cancer deaths per cluster in the primary analyses would be only 0.1, and for the older women it may well be comparably small. Hence, although AgeX is cluster-randomised the comparisons of breast cancer mortality will have virtually the same statistical power as individually randomised comparisons of similar size would have had.

As, however, the numbers excluded for certain particular reasons were substantial, the analyses of these numbers of exclusions (Table 1) had to make appropriate allowance for clustering.

6.7 Additional analyses of all-cause mortality

For both younger and older women analyses of all-cause mortality will be reported. However, they are unlikely to be usefully informative and will not be considered relevant to the interpretation of the primary analyses of breast cancer mortality. For, in both younger and older women, more than 90% of deaths are expected to be from causes other than breast cancer (based on the blinded mortality data to 2024), so there will be negligibly small power for AgeX to assess directly the effect of additional breast screening on all-cause mortality.

6.8 Additional analyses: women not in the main analyses population

Of the 4.5 million women randomised, 3.0 million are in the main analyses population and 1.5 million are not. Analyses of adherence and of outcomes in various categories of those excluded could be of interest and will be reported, but will not be part of the primary and main subsidiary analyses and will not contribute directly to the interpretation of these analyses.

6.9 Safety analyses

Given the statistical difficulty of assessing the effects of screening on mortality from breast cancer, there can be no realistic expectation of demonstrating directly any effects of breast screening on mortality from any other causes, or on all-cause mortality. Nevertheless, safety analyses will be reported that, both in younger and, separately, in older women, report the numbers of deaths from each 3-character ICD category with at least 10 deaths, and of mortality from the aggregate of all causes other than breast cancer. Any P-values ≤ 0.05 generated by these hundreds of analyses will, however, be accompanied by Bonferroni corrections to help allow for the multiplicity of such comparisons.

6.10 Statistical software

SAS version ≥ 9.4, StataNow version ≥1 8.5 and R Studio version ≥ 4.2.3 will be used.