PS 1-32 USING AN INSTRUMENTAL VARIABLE TO DETERMINE THE CAUSAL EFFECT OF OPPORTUNISTIC MAMMOGRAPHY ON BREAST CANCER MORTALITY

Sunday, October 23, 2016
Bayshore Ballroom ABC, Lobby Level (Westin Bayshore Vancouver)
Poster Board # PS 1-32

Mette Lise Lousdal1, Tim L. Lash2, M. Alan Brookhart3, Ivar S. Kristiansen4, Mette Kalager4, Peter Vedsted1 and Henrik StÝvring1, (1)Aarhus University, Department of Public Health, Aarhus, Denmark, (2)Emory University, Rollins School of Public Health and Winship Cancer Institute, Atlanta, GA, (3)University of North Carolina at Chapel Hill, Gillings School of Global Public Health, Chapel Hill, NC, (4)Oslo University, Department of Health Management and Health Economics, Oslo, Norway
Purpose: To estimate the impact of mammography among asymptomatic women on breast cancer mortality by means of an instrument variable: variation in general practices’ mammography referrals as a proxy for mammography preference. 

Method: Closed-cohort study of women aged 50-69 years without prior breast cancer registered with a general practice in 2002-2003 in regions without organized screening identified within the Danish Civil Registration System. Data on mammograms from the Danish National Patient Register were linked with general practices through the Danish National Health Service Register. Women were followed until emigration, death, or Dec 31, 2012. Unit of analysis was practice patient population. Exposure was the percentage of women being referred for mammography in 2003 per practice. Outcomes were the log-transformed breast cancer or all-cause mortality rate measured as the relative change in mortality per 1% increase in mammography referrals. We employed confounded conventional linear regression and potentially unconfounded instrumental two-stage least squares regression with the percentage of referred women in 2002 per practice as instrument. Among practices with no observed breast cancer deaths, “undetected” deaths were multiply imputed using interval regression in [0; 1] based on region, referrals, affiliated women and risk factors.

Result: Among 588 practices with 212,126 registered women 889 breast cancer and 20,687 all-cause deaths had occurred in 2012. The median referral percentage in 2003 was 0.7% (25th; 75th percentiles: 0.4%; 1.3%). A one-unit increase in instrument led to a 0.51% (95% confidence intervals: 0.45%; 0.57%) increase in exposure. Age, use of hormone therapy, and reproductive history did not vary substantially across quartiles of exposure and instrument. Conventional regression found a 6% (95% CI: -5%; 19%) increase in breast cancer mortality associated with a 1% increase in referrals, adjusted for risk factors. Instrumental regression found a corresponding -4% (95% CI: -22%; 19%) decrease. All-cause mortality was not associated with mammography referrals. In sub-analyses, follow-up was limited to 2010 and conventional and instrumental regression found 8% and 6% increases, respectively.

Conclusion: We assume that our instrument identifies referral variation in asymptomatic women, because referral of symptomatic women is independent of preference. When a practice increases the referral proportion 1%, breast cancer mortality among affiliated women decreases -4%. Instrumental variable analysis hinges critically on its assumptions, but has the potential to bypass confounding by severity.