Estimating Population and Health Quantities and their Uncertainty - - PowerPoint PPT Presentation

Estimating Population and Health Quantities and their Uncertainty from Data of Limited Quality

Adrian E. Raftery

University of Washington http://www.stat.washington.edu/raftery

Joint work with Leontine Alkema, Samuel Clark, Patrick Gerland and Mark Wheldon In collaboration with the UN Population Division Supported by NICHD

UN EGM on Strengthening the Demographic Evidence Base for the Post-2015 Development Agenda New York October 6, 2015

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 1

Estimating Population and Health Quantities

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 2

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 3

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 4

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific Population by age and sex

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 5

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific Population by age and sex HIV prevalence

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 6

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific Population by age and sex HIV prevalence

Data:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 7

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific Population by age and sex HIV prevalence

Data:

High quality vital registration and health surveillance data for less than half of countries

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 8

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific Population by age and sex HIV prevalence

Data:

High quality vital registration and health surveillance data for less than half of countries In majority of countries, surveys and censuses only

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 9

Estimating Population and Health Quantities

Goal: Estimate current and past demographic and health quantities and their uncertainty, e.g.

Vital rates (fertility, mortality, migration): summary and age-specific Population by age and sex HIV prevalence

Data:

High quality vital registration and health surveillance data for less than half of countries In majority of countries, surveys and censuses only

Multiple data sources, each with their own issues

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 10

Issues

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 11

Issues

Systematic biases:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 12

Issues

Systematic biases:

Non-representative sampling

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 13

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 14

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 15

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 16

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 17

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata. Why is uncertainty assessment needed?

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 18

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata. Why is uncertainty assessment needed?

general assessment of accuracy: now routine (e.g. UNAIDS, opinion polls, DHS, PMA2020, ACS)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 19

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata. Why is uncertainty assessment needed?

general assessment of accuracy: now routine (e.g. UNAIDS, opinion polls, DHS, PMA2020, ACS) assessing changes and differences between outcomes and expectations

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 20

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata. Why is uncertainty assessment needed?

general assessment of accuracy: now routine (e.g. UNAIDS, opinion polls, DHS, PMA2020, ACS) assessing changes and differences between outcomes and expectations making decisions that avoid risks (e.g. national finance ministries for pension planning, school closures)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 21

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata. Why is uncertainty assessment needed?

general assessment of accuracy: now routine (e.g. UNAIDS, opinion polls, DHS, PMA2020, ACS) assessing changes and differences between outcomes and expectations making decisions that avoid risks (e.g. national finance ministries for pension planning, school closures) Statistics NZ a leader: positive experience (Dunstan, Bryant)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 22

Issues

Systematic biases:

Non-representative sampling Poor geographic coverage Recall bias Undercount (censuses)

Sampling variation: between individuals, between strata. Why is uncertainty assessment needed?

general assessment of accuracy: now routine (e.g. UNAIDS, opinion polls, DHS, PMA2020, ACS) assessing changes and differences between outcomes and expectations making decisions that avoid risks (e.g. national finance ministries for pension planning, school closures) Statistics NZ a leader: positive experience (Dunstan, Bryant) Raftery (2014, arXiv): experiences and types of user

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 23

Bayesian Statistical Modeling

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 24

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q),

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 25

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q),

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 26

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 27

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 28

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 29

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

Information from other countries and expert knowledge can be included through the prior

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 30

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

Information from other countries and expert knowledge can be included through the prior Multiple data sources can be included: If there are m data sources (e.g. different surveys) (Data1, . . . , Datam), the likelihoods are multiplied: p(Data|Q) = p(Data1|Q) × . . . × p(Datam|Q).

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 31

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

Information from other countries and expert knowledge can be included through the prior Multiple data sources can be included: If there are m data sources (e.g. different surveys) (Data1, . . . , Datam), the likelihoods are multiplied: p(Data|Q) = p(Data1|Q) × . . . × p(Datam|Q). Estimates can be made for multiple countries at once, using multinational patterns, by a Bayesian hierarchical model.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 32

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

Information from other countries and expert knowledge can be included through the prior Multiple data sources can be included: If there are m data sources (e.g. different surveys) (Data1, . . . , Datam), the likelihoods are multiplied: p(Data|Q) = p(Data1|Q) × . . . × p(Datam|Q). Estimates can be made for multiple countries at once, using multinational patterns, by a Bayesian hierarchical model.

Now basis for UN population projections

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 33

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

Information from other countries and expert knowledge can be included through the prior Multiple data sources can be included: If there are m data sources (e.g. different surveys) (Data1, . . . , Datam), the likelihoods are multiplied: p(Data|Q) = p(Data1|Q) × . . . × p(Datam|Q). Estimates can be made for multiple countries at once, using multinational patterns, by a Bayesian hierarchical model.

Now basis for UN population projections

Automatically gives uncertainty

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 34

Bayesian Statistical Modeling

Inference about a quantity of interest, Q, summarized by its posterior distribution given all data and evidence: p(Q|Data) ∝ p(Data|Q) × p(Q), i.e. Posterior ∝ Likelihood × Prior Unknown parameters (e.g. bias and measurement error variance of surveys), can be included and estimated. Advantages:

Information from other countries and expert knowledge can be included through the prior Multiple data sources can be included: If there are m data sources (e.g. different surveys) (Data1, . . . , Datam), the likelihoods are multiplied: p(Data|Q) = p(Data1|Q) × . . . × p(Datam|Q). Estimates can be made for multiple countries at once, using multinational patterns, by a Bayesian hierarchical model.

Now basis for UN population projections

Automatically gives uncertainty Complex models can be estimated by Monte Carlo methods

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 35

Estimating HIV Prevalance in Generalized Epidemics

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 36

Estimating HIV Prevalance in Generalized Epidemics

Data:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 37

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 38

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 39

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 40

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 41

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 42

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 43

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 44

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

Bayesian model includes:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 45

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

Bayesian model includes:

Standard SIR epidemic model

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 46

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

Bayesian model includes:

Standard SIR epidemic model Bias in ANC data

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 47

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

Bayesian model includes:

Standard SIR epidemic model Bias in ANC data Measurement error in ANC and DHS data

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 48

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

Bayesian model includes:

Standard SIR epidemic model Bias in ANC data Measurement error in ANC and DHS data

Evaluated by UNAIDS Reference Group

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 49

Estimating HIV Prevalance in Generalized Epidemics

Data:

HIV prevalence at ante-natal clinics:

Frequent measurements = ⇒ Good for trends Unrepresentative Poor geographic coverage

National household surveys (e.g. DHS):

(more) Representative Infrequent (e.g. 0-2 DHS’s in a country).

Bayesian model includes:

Standard SIR epidemic model Bias in ANC data Measurement error in ANC and DHS data

Evaluated by UNAIDS Reference Group

Now used for UNAIDS estimation and projection (EPP/Spectrum software)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 50

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 51

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 52

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 53

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 54

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 55

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts. Expert knowledge about measurement error variances.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 56

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts. Expert knowledge about measurement error variances.

Outputs: Joint posterior distribution of all population quantities of interest.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 57

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts. Expert knowledge about measurement error variances.

Outputs: Joint posterior distribution of all population quantities of interest. Improvements:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 58

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts. Expert knowledge about measurement error variances.

Outputs: Joint posterior distribution of all population quantities of interest. Improvements:

Uncertainty is assessed

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 59

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts. Expert knowledge about measurement error variances.

Outputs: Joint posterior distribution of all population quantities of interest. Improvements:

Uncertainty is assessed All population quantitites are estimated simultaneously: trends and uncertainty are estimated in a demographically consistent way.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 60

Bayesian Population Reconstruction

Wheldon et al., (2010, 2013, 2015)

Bayesian hierarchical model for all population quantities and data Inputs:

Bias-corrected initial estimates of age-specific vital rates, net migration and population counts. Expert knowledge about measurement error variances.

Outputs: Joint posterior distribution of all population quantities of interest. Improvements:

Uncertainty is assessed All population quantitites are estimated simultaneously: trends and uncertainty are estimated in a demographically consistent way. Software: popReconstruct R package

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 61

Laos and New Zealand

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 62

Laos and New Zealand

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 63

Laos and New Zealand Data

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 64

Laos and New Zealand Data

We reconstruct female populations for

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 65

Laos and New Zealand Data

We reconstruct female populations for

Laos, 1985–2005 (increased from 1.8 million to 3.0 million)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 66

Laos and New Zealand Data

We reconstruct female populations for

Laos, 1985–2005 (increased from 1.8 million to 3.0 million) New Zealand, 1961–2006 (increased from 1.2 million to 2.1 million)

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 67

Laos and New Zealand Data

We reconstruct female populations for

Laos, 1985–2005 (increased from 1.8 million to 3.0 million) New Zealand, 1961–2006 (increased from 1.2 million to 2.1 million)

Very different data qualities and demographies.

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 68

Total Fertility Rate

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 69

Total Fertility Rate

• Laos

NZ 2 4 6 8 1985 1990 1995 2000 1960 1970 1980 1990 2000

year TFR (children)

legend

• prior

posterior WPP 2010

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 70

Total Fertility Rate

• Laos

NZ 2 4 6 8 1985 1990 1995 2000 1960 1970 1980 1990 2000

year TFR (children)

legend

• prior

posterior WPP 2010

Average 95% Posterior Interval Half-widths Laos 0.30 NZ 0.03

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 71

Life Expectancy at Birth

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 72

Life Expectancy at Birth

• Laos

NZ 50 55 60 65 70 75 80 1985 1990 1995 20001960 1970 1980 1990 2000

year LEB (years)

legend

• prior

posterior

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 73

Life Expectancy at Birth

• Laos

NZ 50 55 60 65 70 75 80 1985 1990 1995 20001960 1970 1980 1990 2000

year LEB (years)

legend

• prior

posterior

Average 95% Posterior Interval Half-widths Laos 0.80 NZ 0.04

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 74

Age Specific Mortality Rate

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 75

Age Specific Mortality Rate

• Laos (1985)

NZ (1986) 1.0 7.4 54.6 20 40 60 80 20 40 60 80

age mortality rate (per 1000)

legend

• prior

posterior

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 76

Age Specific Mortality Rate

• Laos (1985)

NZ (1986) 1.0 7.4 54.6 20 40 60 80 20 40 60 80

age mortality rate (per 1000)

legend

• prior

posterior

Average 95% Posterior Interval Half-widths Laos 2.0 NZ 0.1

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 77

Total Net Migration

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 78

Total Net Migration

• Laos

NZ −20 −10 10 20 1985 1990 1995 20001960 1970 1980 1990 2000

year migrants (000s per year)

legend

• prior

posterior

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 79

Total Net Migration

• Laos

NZ −20 −10 10 20 1985 1990 1995 20001960 1970 1980 1990 2000

year migrants (000s per year)

legend

• prior

posterior

Average 95% Posterior Interval Half-widths

% of Median Laos 570 NZ 80

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 80

Summary

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 81

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 82

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 83

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues Some success in:

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 84

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues Some success in:

Estimating and projecting HIV prevalence in generalized epidemics

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 85

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues Some success in:

Estimating and projecting HIV prevalence in generalized epidemics Reconstructing past and current population from limited data

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 86

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues Some success in:

Estimating and projecting HIV prevalence in generalized epidemics Reconstructing past and current population from limited data

Require systematic consistent data for as long in the past as possible

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 87

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues Some success in:

Estimating and projecting HIV prevalence in generalized epidemics Reconstructing past and current population from limited data

Require systematic consistent data for as long in the past as possible Papers available at http://www.stat.washington.edu/raftery/Research/soc.html

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 88

Summary

Estimates for demography and health in the majority of countries are based on surveys and censuses from multiple sources with biases and measurement error Bayesian approaches can model and take account of all these issues Some success in:

Estimating and projecting HIV prevalence in generalized epidemics Reconstructing past and current population from limited data

Require systematic consistent data for as long in the past as possible Papers available at http://www.stat.washington.edu/raftery/Research/soc.html Software: popReconstruct R package

UN EGM on Strengthening the Demographic Evidence Base For The Post-2015 Development Agenda, New York, 5-6 October 2015 Session 7. A. Raftery (U. of Washington) – Measuring & communicating uncertainty for estimates 89