Estimating maternal mortality I November 1, 2010 Rafael Lozano Professor of Global Health
May 11, 2015
Estimating maternal mortality I
November 1, 2010
Rafael Lozano
Professor of Global Health
Outline
• Dependent variable
o PMDF to maternal mortality rates
• Model form and covariates
2
Challenges in modeling causes of death
• Data are missing for many years
• Non-sampling error can be large in some settings
• There is marked variation in temporal trends across countries
• The available covariates explain only a moderate component of the variance
What we have
4
What we want
5
How can we get there?
• Statistical models can be used to help explore relationships:
o Identify factors (from the literature) that are likely related to maternal mortality
o Estimate the empirical relationship between those factors and the outcome of interest using a regression model
o Use those empirical relationships to inform our estimates of maternal mortality
6
Outline
• Dependent variable
o PMDF to maternal mortality rates
• Model form and covariates
7
Dependent variable selection
• Dependent variable choices:
o Maternal mortality ratio (MMR)
o Proportion of all deaths due to maternal causes (PMDF)
o Maternal mortality rates
• Dependent variable can either be a summary measure or an age-specific measure
• We model the log of the age-specific maternal mortality rates
8
Choice of dependent variable
• Why model rates rather than the PMDF?
o The PMDF is particularly sensitive to other causes of death
─ For example, in the event of an earthquake, epidemic (such as HIV), or increase in RTIs, the PMDF will be influenced
─ This requires not only modeling maternal mortality, but modeling everything else that explains variation in the PMDF
o At the extremes of the PMDF (close to zero, close to one), models can behave unpredictably
9
Why Model Age-Specific Rates?
• Maternal death rate varies by age, we choose to model age-specific rates to allow for different countries to have different levels and time trends in the maternal death rate.
• Shifts in fertility for example to older ages in some countries influences the age-pattern of maternal mortality.
• Modeling all ages combined forces all countries to have identical patterns over time which is undesirable.
10
Processing input data
11
Why use the PMDF to get to rates?
• We do not calculate the MMR or maternal mortality rate directly from the raw data, but calculate the age-specific (i.e., 15-19, 20-24…45-49) fraction of all deaths in women due to maternal causes (PMDF)
• We then multiply these PMDFs by the all-cause adult mortality estimates discussed earlier, to arrive at the number of maternal deaths
• This allows for the correction of underreporting in the level of maternal mortality as the all-causes mortality “envelopes”
• It also may reduce the effect of recall bias in survey data
12
PMDF to population maternal rates
13
From input data source
PMDF to population maternal deaths, by age
Input data source
Age group # of maternal deaths
# of all-cause deaths
PMDF National mortality
envelopes
Population maternal deaths
15-19 80 1,000 8.0% 1,355 108
20-24 132 1,100 12.0% 1,990 239
25-29 132 1,100 12.0% 3,775 453
30-34 143 1,300 11.0% 4,935 543
35-39 126 1,400 9.0% 5,700 513
40-44 90 1,500 6.0% 6,575 395
45-49 61 1,900 3.2% 7,725 247
14
Outline
• Dependent variable
o PMDF to maternal mortality rates
• Model form and covariates
15
Form of the regression model
• Rates can be modeled directly, as with OLS or robust linear approaches (such as Huber-White, Tukey, or median regression)
o Robust approaches are important because of the presence of outliers, zeros, and other extreme observations
• Rates can also be modeled using count models such as Poisson or negative binomial models
o The data does not meet the Poisson assumption that the mean equals the variance, in other words, the data is over-dispersed
o The degree of over-dispersion is related to age, which can be allowed for using the generalized negative binomial model
16
So what is related to maternal mortality?
• Fertility
• GDP
• Education
• Neonatal mortality
• HIV prevalence
• Coverage of skilled birth attendance or in-facility birth
• Others?
• And what do we have in a complete time series for all countries?
17
Transforming the covariates
• Examine the relationship between each of these covariates and the log of the maternal mortality rate (dependent variable)
• Transformations for model:
o log of the total fertility rate
o log of the distributed lag of GDP per capita
o HIV-squared as well as HIV sero-prevalence
• Look for co-linearity
o SBA highly co-linear with neonatal mortality and GDP per capita, and was also only available for 1986-2008
18
TFR vs. ln(maternal mortality rate)
19
-4-2
02
46
ln(m
ater
nal m
orta
lity
rate
)
0 2 4 6 8TFR
15-19 20-2425-29 30-34
35-39 40-4445-49
First stage regression model
20
Robust Regression Coefficient Std. Error
Intercept 4.715 0.100ln(TFR) 1.903 0.022
ln(GDP per capita) -0.511 0.010Neonatal mortality 13.662 0.721
Education -0.086 0.003HIV 0.108 0.005HIV² -0.001 0.000
Age 15-19 -1.176 0.021Age 20-24 -0.374 0.020Age 25-29 -0.077 0.020Age 35-39 -0.165 0.020Age 40-44 -0.633 0.021Age 45-49 -1.390 0.025
21
22
23
24
25
26
27
28
29
30
31
HIV counterfactual
• HIV is believed to be a major contributor to maternal deaths
• What would happen to maternal mortality if we “turned off” the effect of HIV at the population level?
o Develop a counterfactual scenario: what would have happened with maternal mortality if there had been no HIV
• In the linear model, switch HIV prevalence to zero, rather than its observed value
32