« A case–control study to evaluate the impact of the breast screening programme on breast cancer incidence in England »

Blyuss O., Dibden A., Massat N.J., Parmar D., Cuzick J., Duffy S.W., Sasieni P. - DOI:https://doi.org/10.1002/cam4.5004

The journal Cancer Medicine has just published a new study on quantifying the overdiagnosis associated with breast cancer screening. This study is based on a case-control method with matching and estimation of overdiagnosis by odds ratios (OR) from conditional logistic regression. The method is interesting, but the study has methodological problems that distort the results.

The authors determine the ORs for the year of the last screening mammogram and subsequent years. They then use these ORs to quantify overdiagnosis, not only for the years from which the ORs were calculated
but also for earlier years, including the year of the first screening mammogram. This is only correct if the ORs remain approximately constant from the first mammogram to the last. This is not the case.

In the year of screening, the incidence of breast cancer diagnoses is increased because subclinical cancers are found by screening. Logically, in subsequent years, the incidence is reduced because of cancers
already diagnosed in advance in the year of screening. Due to lead time, this effect lasts for a few years and certainly for at least 3 years (Table 3 in the study suggests that the impact of lead time would
last for about 6 years). With a screening mammogram every 3 years, the effect of the lead time is not complete when the next mammogram is performed. As a result, the incidence of breast cancer diagnoses is lower
for the ith screening mammogram than for the 1st.

This difference between the incidence of cancer at the first screening mammogram and the incidence of cancer at subsequent mammograms is far from marginal and is found in all screening databases.

For England, data from the NHS Breast Screening Programme for 2010-2011 demonstrate an excess of breast cancer diagnoses at the first screening compared to subsequent screenings
(see appendix at the bottom of the page).

For France and the year 2011, the following table is available (taken from "National performance indicators for the breast cancer screening program for the period 2010-2011", downloadable from the Santé
publique France website at
https://www.santepubliquefrance.fr/content/download/53655/file/indicateur-globaux-web-nat-2010-2011.pdf)

The incidence of breast cancer is much higher at the initial screening than at subsequent screenings.
**Thus, there is a methodological error in applying the OR from the year of the last screening mammogram to the year of the 1st screening mammogram**.

The data in the 1st row of the table above suggest that, for women aged 50-54, the incidence is about 1.9 times higher for initial screening than for subsequent screening. Let's apply this factor to correct
the incidence estimates associated with the 1st screening episode at age 50, as shown in Table 4 of the article, an excerpt of which is reproduced below:

The correction is done simply by multiplying the incidence in the screened population, 732.4, by 1.9. The correction leads to an increase in the
incidence in the screened population and consequently in overdiagnosis of 659 cases.

After this correction, the following table can be established:

Erroneous values published in the study |
Corrected values | |

Incidence in the general population (per 100,000 women) | 9 706 | 9 706 |

Incidence in an unscreened population (per 100,000 women) | 9 156 | 9 156 |

Incidence in a screened population (per 100,000 women) | 9 835 | 10 494 |

Overdiagnosis (per 100,000 women) | 679 | 1 338 |

Probability of overdiagnosis for a screened woman | 0.007 (7 chances out of 1000) |
0.013 (13 chances out of 1000 |

Percentage of overdiagnosis among cancers | 7% | 14% |

Percentage of overdiagnosis among cancers diagnosed during screening | 9.5% | 19% |

It can be seen that **the error committed is not negligible** since the correction almost doubles the estimate of overdiagnosis.

The correction made above depends on the value used to determine the difference between the OR of the first screening and the ORs of subsequent screenings. We used a correction factor of 1.9, probably well
adapted to the French situation but whose application to English data may be questionable. We will now use another way of estimating overdiagnosis.

One of the study results is a significant increase in breast cancer diagnoses in women who participated at least once in screening compared with women who never participated, with an OR calculated at 1.22,
with a 95% confidence interval of 1.18 to 1.26. In practical terms, this OR means that women who have participated in screening at least once have a 1.18 to 1.26-fold increase in the probability of being
diagnosed with breast cancer compared to other women. In addition to the fact that this increase in the risk of a diagnosis of breast cancer is far from negligible, it makes it possible to estimate the
proportion of overdiagnosis in cancers diagnosed in a population invited to screening.

Mathematically, the proportion S of overdiagnosis in diagnosed cancers is written as: S = (I_{with} - I_{without}) x D / I_{global}

where I_{with} is the incidence of cancers in women participating in screening, I_{without} is the incidence of cancers in women not participating in screening, I_{global}
is the overall incidence of cancers (in a population mixing participating and non participating women) and D is a correction factor equal to screening participation.

We have 2 additional equations:

OR ≈ RR = I_{with} / I_{without} where OR = odds ratio and RR = relative risk

I_{global} = (1 - D) x I_{without} + D x I_{with}

Combining the above 2 equations, we arrive at: I_{global} = (1 - D) x I_{without} + D x I_{without} x OR

Hence : I_{without} = I_{global} / (1 - D + D x OR) et I_{with} = I_{global} x OR / (1 - D + D x OR)

We deduce: S = (I_{global} x OR / (1 - D + D x OR) - I_{global} / (1 - D + D x OR)) x D / I_{global}

Which simplifies to: S = (OR - 1) x D / (1 - D + D x OR)

With D = 0.7 (70% participation in organized screening) and the bounds of the confidence interval of the OR (1.18 and 1.26), we obtain the following range for S:

low limit: (1.18 - 1) x 0.7 / (0.3 + 0.7 x 1.18) = 0.11 ; high limit: (1.26 - 1) x 0.7 / (0.3 + 0.7 x 1.26) = 0.15

With D = 1 (all women participate in screening), we obtain the following range for S:

low limit: (1.18 - 1) x 1 / (0 + 1 x 1.18) = 0.15 ; high limit: (1.26 - 1) x 1 / (0 + 1 x 1.26) = 0.21

Thus, under standard conditions corresponding to those of the study (invitation every 3 years, 70% participation in screening, and an average of 3 mammograms per participating woman), overdiagnosis would
represent between 11 and 15% of cancers.

In a population of women all participating in organized screening, overdiagnosis would represent between 15 and 21% of all cancers and between 20 and 28% of cancers diagnosed during screening mammography (using
the same 11/8 factor as in the study to account for interval cancers).

**These estimates confirm that overdiagnosis is much more frequent than reported by the study authors**.

This new study aimed at quantifying overdiagnosis contains **a serious methodological error**. This error leads to **a significant
underestimation of overdiagnosis** and **negates any credibility of the study's results and conclusion**.

This study should therefore not be taken into account in estimating the benefit/risk
balance of breast cancer screening, nor should it be included in meta-analyses aimed at quantifying the frequency of overdiagnosis.

Blyuss O., Dibden A., Massat N.J., Parmar D., Cuzick J., Duffy S.W., Sasieni P.

A case–control study to evaluate the impact of the breast screening programme on breast cancer incidence in England

DOI:https://doi.org/10.1002/cam4.5004

Return

Official data on breast cancer screening in England for 2010-2011 can be downloaded from
https://digital.nhs.uk/data-and-information/publications/statistical/breast-screening-programme/breast-screening-programme-england-2010-11.

From these data, it is possible to construct the following tables:

Number of women | Number of cancers | Incidence of cancers (/1000 women) | |

First dépistage | 294,290 | 2,338 | 0.0079 |

Subsequent screening with previous screening less than 5 years old |
1,342,325 | 9,747 | 0.0073 |

Subsequent screening with previous screening older than 5 years |
90,962 | 1,016 | 0.0112 |

Number of women | Number of cancers | Incidence of cancers | |

Women aged 50 to 52 | 287,036 | 1,955 | 0.0068 |

Women aged 53 or 54 | 182,754 | 947 | 0.0052 |

Women aged 55 to 59 ans | 421,096 | 2,533 | 0.0060 |

Women aged 60 to 64 | 452,196 | 3,785 | 0.0084 |

Women aged 65 to 70 | 384,495 | 3,881 | 0.0101 |

50 to 52 | 53 or 54 | 55 to 59 | 60 to 64 | 65 to 70 | |

First screening | 239,412 | 22,366 | 19,122 | 8,763 | 4,637 |

Subsequent screening with previous screening less than 5 years old |
46,641 | 158,149 | 374,372 | 410,359 | 352,804 |

Subsequent screening with previous screening older than 5 years |
983 | 2,239 | 27,612 | 33,074 | 27,054 |

The 2nd table confirms that cancer incidence increases with age. The 3rd table shows that younger women are overrepresented in the first screenings, and older women are overrepresented in the subsequent
screenings. It is, therefore, necessary to adjust for age to see if the incidence of cancers is different for the first and subsequent screenings.

Let us assume that the incidence of cancers remains the
same whether it is a 1st or a subsequent screening. Under this assumption, the incidence depends only on the age group. We can calculate the expected number of cancers by multiplying the number of women in
the age group by the incidence for that age group. This gives the table below:

50 to 52 | 53 or 54 | 55 to 59 | 60 to 64 | 65 to 70 | Total | |

First screening | 239,412 x 0.0068 = 1,631 | 22,366 x 0.0052 = 116 | 19,122 x 0.0060 = 115 | 8,763 x 0.0084 = 73 | 4,637 x 0.0101 = 47 | 1,982 |

Subsequent screening with previous screening < 5 years |
46,641 x 0.0068 = 318 | 158,149 x 0.0052 = 820 | 374,372 x 0.0060 = 2,252 | 410,359 x 0.0084 = 3,435 | 352,804 x 0.0101 = 3,561 | 10,385 |

Subsequent screening with previous screening > 5 years |
983 x 0.0068 = 7 | 2,239 x 0.0052 = 12 | 27,612 x 0.0060 = 166 | 33,074 x 0.0084 = 277 | 27,054 x 0.0101 = 273 | 734 |

We can then compare the observed numbers with the expected numbers under the assumption of no difference between the first and subsequent screening:

Observed number of cancers | Expected number of cancers | Difference | |

First screening | 2,338 | 1,982 | + 356 |

Subsequent screening with previous screening less than 5 years old |
9,747 | 10,385 | - 638 |

Subsequent screening with previous screening older than 5 years |
1,016 | 734 | + 282 |

**The differences between the observed and expected numbers of cancers are statistically significant** (p < 0.00001 with the Chi2 test) **and allow us to reject the hypothesis of no
difference between 1st and subsequent screenings**. More specifically, we confirm that the first screening generate an excess of cancer diagnoses. In contrast, subsequent screenings with a previous
screen less than 5 years old results in a deficit of cancer diagnoses. We also see that when the previous screening is more than 5 years old, the effect of the advance to diagnosis disappears since
we find a surplus of cancer diagnoses.

Return

Dernière mise à jour le 09/08/2022