Estimation and Probabilistic Linkage in Sample Surveys of Anonymous Organisations

<p>Drug use takes on many forms, normally this will be just the occasional alcoholic drink, certain individuals drug use develops into habitual use, or more extreme drugs, and then into full addiction. Some of these addicted individuals realise the harmful nature of their addition and join the anonymous support group, Narcotics Anonymous. This study focus' on the creation of population size estimates, and an estimate of the size of the persistent population between two survey years. These estimates are created from the 2004 and 2008 surveys run by the Narcotics Anonymous Fellowship, as this is an anonymous organisation with no register of the membership database maintained. Population size estimation for an anonymous organisation is established using simulation methods. The bootstrap estimation was used to estimate characteristics about the two populations. Probabilistic matching was used to identify individuals who were in both the 2004, and 2008 surveys. Once identi ed, a logistic regression model was used to establish what impacts an individual to remain in the programme. Factors that impacted an individual being persistent in the population included the individual education, employment status, and if they had worked through all the 12 steps of Narcotics Anonymous.</p>

Estimation and Probabilistic Linkage in Sample Surveys of Anonymous Organisations

<p>Drug use takes on many forms, normally this will be just the occasional alcoholic drink, certain individuals drug use develops into habitual use, or more extreme drugs, and then into full addiction. Some of these addicted individuals realise the harmful nature of their addition and join the anonymous support group, Narcotics Anonymous. This study focus' on the creation of population size estimates, and an estimate of the size of the persistent population between two survey years. These estimates are created from the 2004 and 2008 surveys run by the Narcotics Anonymous Fellowship, as this is an anonymous organisation with no register of the membership database maintained. Population size estimation for an anonymous organisation is established using simulation methods. The bootstrap estimation was used to estimate characteristics about the two populations. Probabilistic matching was used to identify individuals who were in both the 2004, and 2008 surveys. Once identi ed, a logistic regression model was used to establish what impacts an individual to remain in the programme. Factors that impacted an individual being persistent in the population included the individual education, employment status, and if they had worked through all the 12 steps of Narcotics Anonymous.</p>

1369Probabilistic data linkage of health facility and individual data using PMA2020 surveys

Background Most studies rely on clustered analyses to study how the characteristics of health facilities influence individual outcomes. Our aim was to perform a probabilistic linkage between individual and health facility data to enable individual-level analyses. Methods We linked data from the most recent female questionnaire from 11 countries monitored by the Performance Monitoring for Action 2020 to a master health facility dataset (appending all rounds of surveys). Only women that reported which type of facility they visited were considered in the analysis. A probabilistic linkage was performed using 13 blocking variables (e.g., facility type and cluster of residence/location of the woman/facility) and 11 matching variables (e.g., types of contraceptive methods used/offered by the women/facility). Each concordant matching variable received a + 1 score, or a 0 score otherwise. We assessed linkage quality by pooled odds ratio of non-matches according to wealth tertiles (richest vs. poorest) and area of residence (urban vs. rural) using a meta-analytical approach. Results A total of 21,102 women and 7,056 facilities were considered in the linkage process. The average match rate was 57.9%, ranging from 42.5% in Indonesia to 69.1% in Burkina Faso. The pooled odds of non-match were 74% higher for the richest women compared to the poorest, and 67% higher for women living in urban areas compared to rural areas. Conclusions High match rates were achieved in countries with sufficient information on public and private facilities. The lack of information about private facilities contributed to the higher odds of non-match among the better off. Key messages We performed a probabilistic linkage approach to link individual and health facility data, making it possible to understand how the characteristics of health facilities can influence individual-level outcomes. Our findings also bring light to the importance of sampling both public and private facilities, aiming to maximise match rates and reduce differences on match rates according to socio demographic characteristics of the sample.

Reporting Incidences of Neuroblastoma in Various Resource Settings

PURPOSE The incidences of neuroblastoma (NB) differ significantly between various resource settings because of varying quality of cancer registries and underdiagnoses. This study aimed to evaluate current regional variations as reported by international cancer registries and the theoretical and reported differences in international NB incidences and to evaluate South Africa (SA) as a case for variable reporting. METHODS A comprehensive literature review on registries reporting on NB was performed to construct incidence tables. The SEER Program incidence of 10.5/million children was used to calculate the expected number of NB cases for each country. Registry data of NB cases between 2000 and 2016 were requested from The South African National Cancer registry (SA-NCR) and the South African Children's Tumour Registry (SACTR) for comparison and to perform a probabilistic linkage study. RESULTS Internationally, incidences varied between –97.1% and +80% compared with the SEER program. SA under-reported NB cases by an estimated 74.2%. Between 2000 and 2016, the SA-NCR reported between 23 and 51 cases/year, whereas the SACTR reported between 18 and 57 cases/year for the same period. The incidence reported by the SA-NCR varied between 1.5 and 2.8/million children under 15-year per year, whereas the SACTR reported 1.74-2.6 cases/million children. Both registries reported incidences less than high-income country. A probabilistic record linkage of the two registries resulted in a combined incidence of 2.9 cases/million children. CONCLUSION As with most low- and middle-income countries, SA has either a lower incidence or underdiagnoses of NB cases. The reasons for under-reporting are not clear, but can be due to undiagnosed NB cases with spontaneous regression, missed possible cases because of lack of autopsies, and diagnosed cases not recorded in registries.

Design and Creation of Multi-Source Enduring Linked Assets

IntroductionThe Australian Institute of Health and Welfare is collaborating with a range of government and other institutions to build enduring data assets for improving analysis and informed policy outcomes. There were lessons learnt that can be shared, in addition to the architecture and linkage techniques. The enduring assets were created by linking States’ and Territories’ health or welfare data to Commonwealth datasets such as Medicare Consumer Directory (MCD), Residential Aged Care (RAC) and National Death Index (NDI) data. The linkage was carried out by the Australian Institute of Health and Welfare (AIHW) Data Integration Services Centre (DISC). Objectives and ApproachTo create the integrated asset, the linkage spine was assembled by de-duplicating and linking MCD and NDI data. The states’ datasets and other commonwealth datasets involved in the project were linked to this spine. Each unique individual in the spine was assigned Personal Project Number (PPN) which was added to each record linked to the individual. The unlinked individuals from these datasets were de-duplicated and assigned different PPNs. Names, dates of birth and addresses were used in probabilistic linkage process. To enable investigators to interrogate the sequences of the events without releasing the exact dates, the central events file was created. It contains date differences for every event in the asset, calculated as the difference (in days) between event and not released “date zero”, different for each individual. ResultsBetween 96% and 99% of records in the supplied datasets were linked to the spine with linkage accuracy at least 98.5%. The linkage rates depends on the data completeness and the nature of the datasets as not all individuals accessing states’ servicers are eligible for Medicare. Conclusion / ImplicationsThe person-focused de-identified analytical assets allow to study journeys of the individual through Australian health and welfare systems which transcends jurisdictional boundaries.

Abstract 16883: Validation of a Deep Neural Network Electrocardiographic-Age as a Mortality Predictor: The CODE Study

Introduction: Aging affects the electrocardiogram (ECG) with a higher incidence of abnormalities in older patients. ECG-age can be predicted by artificial intelligence (AI) and can be used as a measure of cardiovascular health. Hypothesis: ECG-age predicted by AI is a risk factor for overall mortality. Methods: The Clinical Outcomes in Digital Electrocardiography (CODE) study is a retrospective cohort with a mean follow-up of 3.67 years.The dataset consists of Brazilian patients, mainly from primary care centers. Two established cohorts, ELSA-Brasil, of Brazilian public servants, and SaMi-Trop, of Chagas disease patients, were used for external validation. 2,322,513 ECGs from 1,558,421 patients over 16 years old that underwent an ECG from 2010 to 2017 were included. A deep convolutional neural network was trained in order to predict the age of the patient based solely on ECG 12-lead tracings. The ECG database was split into 85-15% training and test datasets, respectively. Death was ascertained using probabilistic linkage with Brazil′s mortality information data. The Cox regression model, adjusted by age and sex, was used for statistical analysis. The model was validated in two cohorts: ELSA-Brasil (n=14,263) and SaMi-Trop (n=1,631). Results: he mean predicted ECG-age was 52.0 years (±18.7) with a mean absolute error of 8.38 (±7.0) years. Patients with ECG-age >8y older than chronological age had higher mortality rate (HR 1.79, 95%CI 1.69-1.90; p<0.001), whereas those ECG-age >8y younger than chronological age were associated with a lower mortality rate (HR 0.78, 95%CI 0.74-0.83; p<0.001). These results were similar in ELSA-Brasil and SaMi-Trop external validation cohorts (HR 1.75, 95%CI 1.35-2.27; p<0.001;HR 2.42, 95%CI 1.53-3.83; p<0.001 for >8y difference, retrospectively; HR 0.74, 95%CI 0.63-0.88; p<0.001;HR 0.89, 95%CI 0.52-1.54; p=0.68 for <8y difference, respectively). Conclusions: ECG-age, predicted by AI, can be useful as a tool for risk stratification of mortality.

Methods for estimating mis-reporting of TB case notification in high burden settings with weak surveillance infrastructure (Preprint)

BACKGROUND The greatest risk of infectious disease under-notification occurs in settings with limited capacity to reliably detect it. WHO guidance on measurement of mis-reporting is paradoxical, requiring robust, independent systems to assess surveillance completeness. OBJECTIVE Methods are needed to estimate under-notification in settings with weak surveillance systems that do not meet WHO preconditions. This study aims to design tuberculosis (TB) inventory study methods that balance rigor with feasibility for high need settings. METHODS We choose to census most health facilities (HF) and laboratories, restricted reliance upon probability proportional to size sampling to HF types with no capacity to notify. Applying distinct analytical approaches for bacteriologically confirmed versus clinical TB limited the need for extrapolation. At the request of public local health stakeholders, the scope of the TB inventory study methodologies was broadened to include the identification of factors responsible for under-notification and acceptability of potential solutions. RESULTS Retrospective data collection over longer time horizons minimizes bias due to seasonality and measures “natural” recording and reporting behaviors. Leveraging a priori knowledge, minimizing recourse to inference, manual entry, use of transparent probabilistic linkage methods, incentivizing private sector participation, and cross-border case verification help to generate valid estimates despite challenging conditions. CONCLUSIONS Adaptive study designs permit rigorous, relevant, ethical inventory studies in the countries that need them even in the absence of WHO established preconditions. Use of triangulation techniques, minimizing recourse to extrapolation, and a strategic focus on the practical needs of local stakeholders, yielded reasonable misreporting estimates and, crucially, viable policy recommendations.