An Integrated U.S. National Mortality Database by Immigration status - Promises and Issues

Transcription

1 An Integrated U.S. National Mortality Database by Immigration status - Promises and Issues Mandi Yu 1, Joe Zou 2, Benmei Liu 1, and Eric (Rocky) Feuer 1 1 : National Cancer Institute; 2 : Information Management Services, Inc. NAACCR 2018

2 Introduction Considerable growth in size and share of immigrants in the U.S.: 9.6 million (4.7%) in 1970 to 42.4 million (13.3%) in 2014 Increasing interest to investigate and monitor disparities in cancer incidence and mortality by immigration status, which is often operatized as nativity (US-born vs. Foreign-born) Most studies have focused on data drawn from special studies, or limited to local or regional data Cancer research at the national level is scarce, partly due to lack of consolidated database and analytic tools

3 Objectives Create a SEER*Stat mortality database to support calculation of rates by nativity Extend the dimension of routine population estimates by including nativity using samples from the Census Long Form Survey and American Community Survey Link with NVSS death count data to create the SEER*Stat mortality database Evaluate the precision of population estimates Examine the validity of using the current variance estimation method for making inference about mortality rates by nativity Database release plan and future research areas

4 Background on Age-Adjusted Rate Numerator xx: observed number of deaths (or diagnosed cancer cases for calculating incidence rate) Denominator nn: population at risk JJ Age-standardized rate: yy = jj=1 ff jj xx jj nn jj ff jj is the standard population share for age group jj yy jj = xx jj nn jj is the rate for age group jj (age-specific)

5 Variance and Interval Estimation Current method (Tiwari et al. 2006) assumes fixed nn jj, since estimates are based on decennial censuses: JJ vaaaa yy = jj=1 ff jj 2 xx jj nn jj 2 with a modified Gamma interval In the Nativity setting, nn jj ~NN μμ jj, σσ jj 2 because nn jj has to be estimated from survey samples, where σσ jj 2 is the variation in nn jj due to sampling error and non-sampling survey error Therefore, if directly applied, the current method may lead to Underestimated variance + under-covered intervals Possibility of false positive findings

6 Data Source-Numerator U.S. death counts from the NVSS Stratified by 19 age group (<0, 1-4,,85+), gender, race/ethnicity (NH-White, NH-Black, NH-API, NH-AIAN, and Hispanic), nativity (US-born and Foreign-born), and 50 U.S states plus D.C. Nativity is highly complete with <1% of missing data Except that Georgia experienced unusual high incomplete data with 60%~97% of missing data in 2008 and 2009 due to delay in switch to the new format of death certificate files

7 Data Sources-Denominator 2000 census long-form survey sample (~16%) annual 1% ACS IPUMS-USA samples ( with bridged single-race variable Missing data imputed using chained equations method to overcome under-estimate: race/ethnicity (missing ~ %) and nativity (missing <0.01%) Estimated population totals and standard errors accounting for complex sampling features

8 SEER*Stat Incidence Database Linked numerator and denominator and created a SEER*Stat time-series incidence database SEER*Stat supports the calculation of age-adjusted rates Variance and confidence intervals are based on Tiwari method without considering errors in the denominator data

9 Snapshots of SEER*Stat Database 1. Select Database 2. Session Setup 3. Output: Mortality Rate Matrix

10 Illustrative Example: Age-Adjusted All-Cancer Mortality Rates by Race/ethnicity and Nativity, U.S., Several observations: US-born Blacks have the highest rate, second by Whites (regardless of birthplace) Rates are lowest among foreign-born Asians and Hispanics Within race/ethnicity, nativity disparity is greatest for Blacks and smallest for Asians

11 This image cannot currently be displayed. Precision of Population Estimates Coefficient of Variation (CV) of Population Estimate by Sex, Age, Race and Nativity Here, CV is a precision measure, and CV=Standard Error/Point Estimate State Level U.S. Level

12 Monte-Carlo Study To examine the impact of the error in the denominator on variance and interval estimation of mortality rates We incorporate the error into variance estimation through the use of Monte-Carlo simulations Then we compare them with those obtained using standard method without considering the additional error

13 Simulation Steps Steps to obtain the coverage rate with errors incorporated Step 1. generate xx ii ~PPPPPPPP cc ii, where cc ii is observed 2013 death counts Step 2. generate nn ii ~TTTT uu ii, σσ ii 2, aa = 0, bb = +, where uu ii and σσ ii 2 are estimated population and variance from 2013 ACS sample Step 3. calculate age-adjusted rate, variance, and confidence interval (CI) using the Tiwari method, and determine whether CI covers the JJ population estimand Y = jj=1 ff jj cc jj μμ jj Step 4. repeat Step ,000 times and calculate proportion of times that the CI covers the population estimand Repeat the same process but remove Step 2 gives the coverage with the errors ignored

14 Real data scenarios Cancer site (from common to rare) Female: Lung, Breast, CRC, Cervix, and Vagina Male: Lung and CRC Demographic/geographic class Race/ethnicity (NH-White, NH-Black, NH-AIAN, NH-API, Hispanic) Nativity (US-born, Foreign-born) Geographic area (US, 50 States plus D.C.)

15 Coverage Rate by Age Group Note: In each plot, dots represent cancer site, gender, race/ethnicity and state combinations

16 Coverage Rate for Age-standardized Rate Conservative CI under current Tiwari method Under-coverage

17 Conservative for small counts

18 Coverage Rate by Population Estimate CV If the CV can be controlled to be <0.05, the majority of coverage is greater than 80%, and most are over 90%.

19 Issues Monte-Carlo (MC) simulations suggest that failure to incorporate the error may seriously under-estimate the variance Unstable denominators are more likely to occur with fine stratifications- aggregation improves precision Preliminary evaluation of Monte-Carlo CIs (not shown) suggest potential numeric convergence problem because the joint distribution of numerator and denominator was not considered In addition, MC method is computationally expensive to implement in SEER*Stat A close form variance and interval estimators are needed

20 Data Release Plan Aggregate geography to the U.S. or U.S. regions Aggregate multiple years, e.g. 3-year interval with , , and Explore possibility to expand nativity to sub-hispanic or sub- Asian categories Need to ensure the cv of the population estimate is 0.05 or less