Minor influence of self-selection bias on the effectiveness of breast cancer screening in case-control studies in the Netherlands

Abstract

Background

Self-selection bias is considered to be a problem when evaluating the effectiveness of breast cancer service screening in case-control studies.

Objective

Using the incidence-based mortality method (IBM), a correction factor for the potential influence of self-selection can be derived from a group of non-screened women and a group of not-invited women.

Methods

Breast cancer patients, diagnosed in 1990–1995 and between the ages of 50 to 70, were selected from the Netherlands Cancer Registry and five screening regions in the Netherlands. Person-years were calculated for non-screened and not-invited women by using population data available on the number of women invited, the number of women screened and the total population. Incidence-based breast cancer mortality rates according to screening status were calculated for the five screening regions.

Results

Between 1990–1995, 15,541 patients were diagnosed with breast cancer. An analysis of ten-year follow-up after diagnosis resulted in 3903 breast cancer deaths, of which 2631 were not-invited and 345 were non-screened. Poisson regression analysis showed heterogeneity between the regions with a range of IBM ratios from 0.64 (95% CI: 0.46–0.90) to 1.08 (95% CI: 0.82–1.43).

Conclusions

Heterogeneity between the regions stresses the importance of a country- and/or region-specific estimate of self-selection. Adjusting for self-selection bias in the regional case-control studies would not change the breast cancer mortality reduction in three regions and would result in an even larger effect in two regions. Looking at the range of IBM ratios the overall influence of self-selection in the Netherlands is minor.

Introduction

The case-control study is regularly used as a tool for the evaluation of the effect of screening on breast cancer mortality.^1–8 A problem associated with case-control studies is the possible influence of self-selection bias on the magnitude of the effect of screening. Participation in the screening programme is voluntary. It is generally believed that women who decide to participate in screening have a different baseline risk of breast cancer mortality to those who decide not to participate. This difference in baseline risk, expressed as a ratio, can be used to adjust the screening estimate from a case-control study (odds ratio [OR]).⁹

Results from the literature on the direction of self-selection are, however, inconsistent. Friedman and Dubin found that women who refused screening were at lower baseline risk for breast cancer death compared with a control group. Adjustment for this difference in baseline risk causes a larger screening effect (OR further from 1.0) In contrast, Moss et al. found the opposite, which would result in a smaller screening effect (OR closer to 1.0).^10–12

It is to be expected that the factors underlying self-selection may be different in each country or even region that has implemented service screening. However, several studies have applied a correction factor based on estimates calculated from data from the Swedish and Canadian randomized controlled trials (RCT) by Duffy et al. instead of a country-specific or regionally determined selection factor.^3,13 Based on the study by Duffy et al., we anticipated a higher breast cancer mortality rate in the non-screened women. Therefore, the correction factor for self-selection has been estimated as the ratio of the IBM rate in the non-screened women and the IBM rate of the not-invited women.⁹ This ratio can then be applied to a formula to adjust the effect of breast cancer service screening on breast cancer mortality in a case-control study.^9,14

When nationwide screening is offered to all women in the target group, a well-defined control group is no longer available. In this situation, the only period from which a control group can be created is directly before the start of the implementation of a screening programme or during the implementation of the screening programme, when part of the target group is still not-invited. The purpose of this study is to estimate a correction factor for self-selection bias from the implementation period of the service screening programme in five regions of the Netherlands, using the incidence-based mortality method (IBM).¹⁵

Methods

Service screening in the Netherlands started in 1989, initially inviting women in the age group 50–69 biennially, and was fully implemented in 1997. In this period nine screening organizations were responsible for the organization of the screening programme. In this study we included five of the regions, namely Stichting Kankerpreventie en – screening Limburg (SKsL), Stichting Kankerpreventie IKA (SKP IKA), Stichting Bevolkingsonderzoek Noord-Nederland (BBNN), Stichting Vroege Opsporing Kanker Oost-Nederland (SVOKON) and Stichting Bevolkingsonderzoek Borstkanker Zuidwest Nederland (SBBZWN). The regions were chosen based on their differences in urban and rural areas, and early and late implementation of the screening programme. These regions included 57% of the breast cancer patients in the age group 50–69 registered in the Netherlands Cancer Registry in the years 1990–1995.

An incidence-based mortality (IBM) rate can be calculated by dividing the number of breast cancer deaths resulting from incident breast cancer cases diagnosed in a specific period with the accompanying number of person-years. The calendar period for calculation of the IBM rate included the years 1990–1995. During the implementation phase of the Dutch screening programme, the female population in the targeted age group shifted from an uninvited population to an invited population, providing us with contemporaneous groups of not-invited and invited-but-not-screened women.

In order to define the numerator of the breast cancer mortality rates, data of breast cancer patients aged 50–69 and diagnosed in the years 1990–1995 were collected from the Netherlands Cancer Registry. The patient data were linked to the screening registries of the five screening organizations to collect information on screening history and vital status. The data were also linked with the Cause of Death Registry held by Statistics Netherlands to obtain cause of death. The breast cancer patients were divided into three groups according to their screening status: not invited for screening before diagnosis (not-invited), invited and screened before diagnosis (screened), invited and not screened before diagnosis (non-screened). Breast cancer patients were defined as not screened after not participating in the most recent invitation to screening before diagnosis. We followed the breast cancer patients for breast cancer death for a maximum of 10 years after date of diagnosis. Screened patients were not followed since this group does not contribute to estimation of the correction factor.

In order to define the denominator of the breast cancer mortality rates, we calculated the person-years for the period 1990–1995 using aggregated data on the total number of invited women, the total number of screened women derived from the screening organizations, and the age-specific total population derived from Statistics Netherlands. In order to divide the person-years in a group of not-invited and a group of non-screened person-years, we calculated the number of person-years during the implementation period of the screening in each municipality separately. The municipality is the smallest possible unit of which the starting date of screening was known. The population data and starting date of screening were determined and used in a previous study by Otto et al.¹⁶ Further details on the calculation of person-years along with an example can be found in the Appendix.

Results

In total, 15,541 breast cancers were diagnosed in women aged 50–69 in the period 1990–1995. The complete number of person-years for this period was 5,199,451. Of the 8979 not-invited breast cancer patients, 2631 died of breast cancer, whereas of the 968 not-screened patients, 345 died of breast cancer. Table 1 shows patient characteristics of not-invited and non-screened women. The distributions of age at diagnosis (χ² = 5.03, df = 3, P = 0.17) and stage at diagnosis (χ² = 0.68, df = 1, P = 0.41) were comparable between both groups.

Table 1

Age at diagnosis and stage distribution for breast cancer patients divided into non-screened and not-invited patients

Breast cancer patients	Non-screened (%)	Not-invited (%)
Total	968 (100)	8979 (100)
Age at diagnosis
50–54	236 (24.4)	2475 (27.6)
55–59	226 (23.4)	1948 (21.7)
60–64	225 (23.2)	2082 (23.2)
65–69	281 (29.0)	2474 (27.5)
Stage^*
0–1	296 (32.2)	2847 (33.6)
2+	623 (67.8)	5637 (66.4)

For both groups 5% missing

Data were available from five screening regions in the Netherlands. Poisson regression was used to test heterogeneity between the regions. The model which included the interaction terms of the regions fitted significantly better than the model without the interaction terms of the regions (χ² = 0.03), which points towards a difference in the degree of self-selection in the five regions.

Table 2 shows an example of the calculation of the IBM ratio in one region, namely the SVOKON region. The IBM rate in the non-screened women, 9.6 per 10,000 person-years, is somewhat higher than the IBM rate for the not-invited women, 8.8 per 10,000 person-years. Dividing those two IBM rates results in an IBM-ratio of non-screened versus not-invited of 1.08 (95% CI: 0.85–1.37), indicating no self-selection in the SVOKON region.

Table 2

Example of calculation of the incidence-based mortality rates for non-screened and not-invited women in the region of SVOKON

SVOKON Period 1990–1995
Non-screened
Incident cases	271
Breast cancer deaths	99
Person-years	103,358
IBM/10,000 PY	9.6
Not-invited
Incident cases	800
Breast cancer deaths	216
Person-years	244,204
IBM/10,000 PY	8.8

Rate ratio^* (9.6/8.8)	1.08
(95% CI)	(0.85–1.37)

Correction factor for self-selection, estimated by IBM rate non-screened/IBM rate not-invited

In Table 3, the correction factors and their 95% confidence intervals of all regions are given. As for the SVOKON region, the SBBZWN and SKsL region showed no difference in breast cancer mortality risk between the non-screened and not-invited women. The IBM rate ratio in the BBNN and SKP IKA region is lower than one, indicating a lower breast cancer mortality risk in the non-screened women.

Table 3

Correction factors for self-selection bias for five screening regions in the Netherlands

	Breast cancer deaths/person years
Region	Non-screened	Not-invited	Rate ratio (95% CI)
BBNN	36/73,412	702/917,668	0.64 (0.46–0.90)
SKP IKA	117/179,463	798/937,692	0.77 (0.63–0.93)
SKsL	39/63,341	189/282,716	0.92 (0.65–1.30)
SBBZWN	54/68,867	726/1,003,423	1.08 (0.82–1.43)
SVOKON	99/103,358	216/244,204	1.08 (0.85–1.37)

Discussion

The heterogeneity between the regions showed that even in a centrally organized screening programme regional differences can be found. Three of the five regions, namely the SBBZWN, SKsL and SVOKON regions, showed no difference in baseline risk in breast cancer mortality in women not participating to the Dutch screening programme compared with a not-invited population. In the BBNN and SKP IKA regions IBM ratios below 1.0 were found, representing a lower baseline breast cancer mortality risk in non-screened to not-invited women. These ratios can be used as regional correction factors for self-selection bias for the odds ratio estimated in a case-control study to evaluate the impact of breast cancer screening on breast cancer mortality. Without correction in these two regions, the estimated breast cancer mortality reduction is slightly underestimated.

Looking at the magnitude of the correction factors and their confidence interval, heterogeneity is most likely a result of the 0.64 (95% CI: 0.46–0.90) in the BBNN region. Reasons why this region is different from the other regions are not known. An explanation may be the higher attendance rate (83.9% compared with 78.9% nationally¹⁷) and a difference in underlying background risks for breast cancer mortality; it is the most rural area in this study. Furthermore, the use of aggregated person years could have influenced the heterogeneity. Between 1990 and 1995 many municipal borders were redrawn and municipalities renamed.¹⁶ The precise division of the municipalities in the regions in this period is quite difficult to trace and some (differential) misclassification could have occurred.

The reason for a somewhat lower rate ratio in the BBNN and SKP IKA regions is not known. Studies that explored reasons for non-attendance in the Netherlands mentioned mammograms taken outside the service screening programme, practical reasons, pain, no breast cancer in the family and breast self-examination as reasons for not accepting an invitation to screening.^18,19 The possible association between these factors and a lower breast cancer mortality in non-screened women needs further exploration. A population survey performed in Australia also suggested that screening participants would have a slightly higher background risk of breast cancer than non-screened women. The participants were more likely to report a family history of breast cancer a history of breast surgery, and previous use of hormone replacement therapy.⁷ Using a proxy for self-selection bias like differences in risk factors or in incidence of breast cancer is a way of coping with self-selection bias. However, caution is needed when interpreting the results. For example, it has been shown that equal breast cancer incidence in non-screened women and a control group does not necessarily lead to equal breast cancer mortality.¹²

In the region of SKsL we calculated a correction factor of 0.92 (95% CI: 0.65–1.30) in this study. In a previous case-control study in the same region, we calculated a correction factor of 0.84 (95% CI: 0.58–1.21).²⁰ There are two factors that might be responsible for this small change. First, we used a slightly different method to calculate the IBM-ratio. Instead of using the date of diagnosis of each patient as the start of follow-up, we had a start of follow-up time independent from date of diagnosis in the previous study. Second, a small difference in the number of breast cancer deaths retrieved through linkage with the Cause of Death Registry of Statistics Netherlands could also be an influence. However, correction with these factors does not lead to a difference in the impact of the breast cancer mortality reduction by breast cancer screening.

Our correction factors for self-selection calculated for the regions in the Netherlands differ from those used in other case-control studies. Duffy et al. calculated a correction factor for self-selection based on the data from RCTs of mammographic screening.⁹ This correction factor was 1.36, which represents a higher breast cancer mortality in the non-screened compared with the uninvited control group, and indicates a higher baseline mortality rate in the non-screened women compared with the screened women. The use of this correction factor is only justified when the relative mortality of the non-screened population and the population of uninvited women is the same in the programme at issue as observed in the previously published RCTs.⁹ In recent case-control studies, the correction factor of 1.36 has been used to correct the odds ratio for self-selection bias.^3,13 Puliti et al. and Gabe et al. estimated regional or country-specific correction factors of 1.11 and 1.17, respectively, to correct the odds ratio.^4,6 In line with their results, our study demonstrates that 1.36 cannot be interpreted as a constant correction factor. Rather a country-specific or regionally determined correction factor should be used.

The application of date of diagnosis of each patient as start of follow-up was possible because we included non-screened and not-invited women only. When using a before-and-after the start of screening comparison for the effectiveness of screening, a date independent of diagnosis should be chosen to correct for possible lead time bias. Correction for lead time bias is only necessary if screened women are included in the comparison.^14,21,22

Based on our study, the IBM method is an effective tool for calculating the correction factor for self-selection bias in the Netherlands. This is supported by a Swedish study, where they also calculated regional specific correction factors for self-selection bias using the incidence-based mortality method.²³ Our correction factor was based on breast cancer patients diagnosed in the period 1990 to 1995. Correction for self-selection bias of an estimated odds ratio for recent years would assume that the differential between the non-screened and not-invited women does not change over time. This assumption is inevitable, but is supported by the high and stable attendance rate in the Netherlands, which suggests that the non-screened women have not changed over time.

In conclusion, the heterogeneity between the regions shows the necessity for calculating regional or country-specific correction factors, even in a centrally organized screening programme. Looking at the range of IBM ratios in the five screening regions, the odds ratio estimated in a nationally conducted case-control study will show no change or a small increase in breast cancer mortality reduction after correction for self-selection. This indicates that self-selection is only a minor problem in the Netherlands.

Footnotes

Acknowledgements

This work was supported by a grant from the Dutch Cancer Society (project number: KUN 2006–3571). The funding source had no involvement in the study.

We gratefully acknowledge Stichting Kankerpreventie en – screening Limburg (SKsL), Stichting Kankerpreventie IKA (SKP IKA), Stichting Bevolkingsonderzoek Noord-Nederland (BBNN), Stichting Vroege Opsporing Kanker Oost-Nederland (SVOKON), Stichting Bevolkingsonderzoek Borstkanker Zuidwest Nederland (SBBZWN), Netherlands Cancer Registry and Statistics Netherlands for providing the data. We thank Dr. S Otto, Erasmus MC, University Medical Centre, Rotterdam, for providing the population data and the starting dates of screening for the different municipalities.

Calculation of person-years

The calculated number of person-years were based on the number of invited women, the number of screened women and the age-targeted female population for each municipality in the five regions separately. The not-invited number of person-years were calculated by subtracting the number of invited person-years from the total person-years of the targeted population.

The person-years of invited women were estimated based on the number of invited women:

–

year of the start of screening: half of the number of invited women

–

year + 1: all invited women in the year of start of screening and half of the invited women in year + 1

–

year + 2: 95% of the total population of that year, assuming that 5% of the eligible population is not-invited

Adding up the number of invited person-years for every municipality for the total period of 1990–1995 resulted in the total number of invited person-years.

In addition, the non-screened person-years were calculated by subtracting the screened person-years from the total number of invited person-years.

The number of person-years of the screened women were based on the number of screened women:

–

year of the start of screening: half of the number of screened women

–

year + 1: all screened women in the year of start of screening and half of the screened women in year + 1

–

year + 2: half of the screened women in the year of start of screening, all screened women in year + 1 and half of the screened women in year + 2

Adding up the number of screened person-years for every municipality for the period 1990–1995 resulted in the total number of screened person-years.

An example is given below for one municipality (Table A and Table B).

References

Broeders

MJM

, Verbeek

ALM

, Straatman

Repeated mammographic screening reduces breast cancer mortality along the continuum of age.

J Med Screen 2002; 9: 163–7.

Elmore

, Reisch

, Barton

MB.

Efficacy of breast cancer screening in the community according to risk level.

J Natl Cancer Inst 2005; 97: 1035–43.

Fielder

, Warwick

, Brook

A case-control study to estimate the impact on breast cancer death of the breast screening programme in Wales.

J Med Screen 2004; 11: 194–8.

Gabe

, Tryggvadottir

, Sigfusson

, Olafsdottir

, Sigurdsson

, Duffy

SW.

A case-control study to estimate the impact of the Icelandic population-based mammography screening program on breast cancer death.

Acta Radiol 2007; 48: 948–55.

Palli

, del Turco

, Buiatti

A case-control study of the efficacy of a non-randomized breast cancer screening program in Florence (Italy).

Int J Cancer 1986; 38: 501–4.

Puliti

, Miccinesi

, Collina

Effectiveness of service screening: a case-control study to assess breast cancer mortality reduction.

Br J Cancer 2008; 99: 423–7.

Roder

, Houssami

, Farshid

Population screening and intensity of screening are associated with reduced breast cancer mortality: evidence of efficacy of mammography screening in Australia.

Breast Cancer Res Treat 2008; 108: 409–16.

Verbeek

ALM

, Hendriks

, Holland

, Mravunac

, Sturmans

, Day

NE.

Reduction of breast cancer mortality through mass screening with modern mammography. First results of the Nijmegen project, 1975-1981.

Lancet 1984; 1: 1222–4.

Duffy

, Cuzick

, Tabár

Correcting for non-compliance bias in case-control studies to evaluate cancer screening programmes.

Appl Stat 2002; 51 Part 2: 235–43.

10.

Cronin

, Weed

, Connor

, Prorok

PC.

Case-control studies of cancer screening: theory and practice.

J Natl Cancer Inst 1998; 90: 498–504.

11.

Friedman

, Dubin

Case-control evaluation of breast cancer screening efficacy.

Am J Epidemiol 1991; 133: 974–84.

12.

Moss

, Summerley

, Thomas

, Ellman

, Chamberlain

JOP.

A case-control evaluation of the effect of breast cancer screening in the United Kingdom trial of early detection of breast cancer.

J Epidemiol Community Health 1992; 46: 362–4.

13.

Allgood

, Warwick

, Warren

, Day

, Duffy

SW.

A case-control study of the impact of the East Anglian breast screening programme on breast cancer mortality.

Br J Cancer 2008; 98: 206–9.

14.

Duffy

, Tabar

, Chen

HH.

The impact of organized mammography service screening on breast carcinoma mortality in seven Swedish counties.

Cancer 2002; 95: 458–69.

15.

Paci

, Duffy

, Giorgi

Quantification of the effect of mammographic screening on fatal breast cancers: the Florence Programme 1990-96.

Br J Cancer 2002; 87: 65–9.

16.

Otto

, Fracheboud

, Looman

CW.

Initiation of population-based mammography screening in Dutch municipalities and effect on breast-cancer mortality: a systematic review.

Lancet 2003; 361: 1411–7.

17.

Fracheboud

, Otto

, Groenewoud

JH.

National evaluation of breast cancer screening in the Netherlands (in Dutch). IX. Rotterdam: Department of Public Health, Erasmus MC, University Medical Center Rotterdam; 2001

18.

Geel

Redenen en kenmerken niet-deelnemers aan bevolkingsonderzoek borstkanker (in Dutch, internship report). Amsterdam: InHolland; 2009

19.

ten Voorde

Non-respondentenonderzoek. Wat zijn de oorzaken van de non-respons bij het bevolkingsonderzoek naar borstkanker (in Dutch). Stichting Vroege Opsporing Borstkanker; Bevolkingsonderzoek Borstkanker Noord-Nederland; 2008

20.

Paap

, Holland

, den Heeten

GJ.

A remarkable reduction of breast cancer deaths in screened versus unscreened women: a case-referent study.

Cancer Causes Control 2010; 21: 1569–73.

21.

Baker

, Kramer

, Prorok

PC.

Comparing breast cancer mortality rates before-and-after a change in availability of screening in different regions: extension of the paired availability design.

BMC Med Res Methodol 2004; 4: 12–20.

22.

Jonsson

, Nyström

, Tornberg

, Lenner

Service screening with mammography of women aged 50-69 years in Sweden: effects on mortality from breast cancer.

J Med Screen 2001; 8: 152–60.

23.

Swedish Organised Service Screening Evaluation Group.

Reduction in breast cancer mortality from the organised service screening with mammography: 2. Validation with alternative analytic methods.

Cancer Epidemiol Biomarkers Prev 2006; 15: 52–6.