Quantifying Child Mortality Reductions Related to Measles Vaccination

Background This study characterizes the historical relationship between coverage of measles containing vaccines (MCV) and mortality in children under 5 years, with a view toward ongoing global efforts to reduce child mortality. Methodology/Principal Findings Using country-level, longitudinal panel data, from 44 countries over the period 1960–2005, we analyzed the relationship between MCV coverage and measles mortality with (1) logistic regressions for no measles deaths in a country-year, and (2) linear regressions for the logarithm of the measles death rate. All regressions allowed a flexible, non-linear relationship between coverage and mortality. Covariates included birth rate, death rates from other causes, percent living in urban areas, population density, per-capita GDP, use of the two-dose MCV, year, and mortality coding system. Regressions used lagged covariates, country fixed effects, and robust standard errors clustered by country. The likelihood of no measles deaths increased nonlinearly with higher MCV coverage (ORs: 13.8 [1.6–122.7] for 80–89% to 40.7 [3.2–517.6] for ≥95%), compared to pre-vaccination risk levels. Measles death rates declined nonlinearly with higher MCV coverage, with benefits accruing more slowly above 90% coverage. Compared to no coverage, predicted average reductions in death rates were −79% at 70% coverage, −93% at 90%, and −95% at 95%. Conclusions/Significance 40 years of experience with MCV vaccination suggests that extremely high levels of vaccination coverage are needed to produce sharp reductions in measles deaths. Achieving sustainable benefits likely requires a combination of extended vaccine programs and supplementary vaccine efforts.


Supporting Information S1. Data description and coding
Data used in the analyses were derived from multiple sources (Table S1A). Measles deaths were identified from the WHO Mortality Database based on ICD codes shown in Table S1B. Measles vaccine coverage estimates for years prior to 1980 were derived from a variety of country-specific reports shown in Table S1C. The countries, numbers of observations, and observation years with data present for all variables are shown in Table S1D.  Table S1B. International Classification of Disease (ICD) codes used to identify measles deaths* ICD-7  085  Measles  ICD-8  055  Measles  ICD-9 055 Measles ICD-10 B05 Measles ICD-10 B050 Measles complicated by encephalitis ICD-10 B051 Measles complicated by meningitis ICD-10 B052 Measles complicated by pneumonia ICD-10 B053 Measles complicated by otitis media ICD-10 B054 Measles with intestinal complications ICD-10 B058 Measles with other complications ICD-10 B059 Measles without complication * Summary codes corresponding to these detailed list numbers were used to extract country, year, and genderspecific mortality counts for measles-related deaths.

Logarithmic transformation of measles death rates
For analyses of reductions in measles death rates, the rates were log transformed (natural log). For country-years with zero observed measles deaths, the log-transformed rate is undefined. To prevent these observations from being dropped from the analysis, we therefore replaced zero values with the minimum observed rate divided by 10 and then log-transformed all rates. The minimum non-zero observed rate in the dataset occurred in the United States in 1992 and was 5x10 -3 measles deaths per 100,000 children aged 0 through 5. Sensitivity analyses relating to the treatment of zeros and the analysis of measles specific death rates are presented in Supporting Information S6.

Categorizing measles vaccination coverage levels
Measles-containing vaccine (MCV) coverage is a continuous variable running from 0 to 100%. Because the relationship between MCV coverage and measles death rates may be non-linear and because we did not wish to impose a functional form, we categorized MCV coverage into a number of discrete levels. We defined these divisions prospectively so that the number of observations in each level above 0% MCV coverage was nearly equal and cutoffs were divisible by 5. MCV coverage was categorized into the following levels: 0%; 1-59%; 60-79%; 80-89%; 90-94%; and >=95% coverage. We also constructed restricted cubic splines for MCV coverage with knots placed at the same cutoffs

Supporting Information S3. Sensitivity analysis: model fit and specifications
We compared alternative model specifications using both the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). In general a reduction in the AIC or BIC from one model specification to another of approximately 2 or greater indicates a significant improvement in model fit even after penalization for specifications that include additional parameters.
While the dependent variable is measles mortality in children under 5, the base case uses MCV coverage in 12-24 month-olds lagged by 1 year. It is therefore possible that vaccination coverage in prior years (i.e., vaccinated 2 year-olds that are now 4 year-olds) have an effect on under-5 mortality as well. At the same time, if case fatality from measles is higher in younger children, including coverage levels for somewhat older children may attenuate the estimated relationship. We use lagged 5-year average MCV coverage in an alternate model specification. In fact, we considered lags in two ways. First, we constructed averages based on all observations within a 5-year range. For example, if data were only available for periods 2 years and 4 years prior, then only these two observations were used to construct the lagged average. Second, we constructed averages requiring that all MCV coverage values be present in the previous 5-year range. The first approach preserves sample size but makes the exact definition of the lagged average harder to interpret. The second approach maintains a clear definition of the lagged average, but loses sample size, and selects certain countries with longer and more continuous data series (see Table S1D above). Compared to the base-case specification, the main findings of the regression were similar in the alternative specifications, although the magnitude of the impact of MCV coverage above 80% attenuated under the alternatives (Table S3D). Computing percentage reductions in mortality under different coverage levels as above, we found that the two alternative 5-year lagged specifications predicted reductions in measles deaths at MCV coverage of 85-94% of 87% or 88% compared to 92% in the base case specification.

Supporting Information S4. Sensitivity analysis: alternative regressions among country subsets defined by income or region
We estimated our main model specification on subsets of our data, confining our analysis to countries with higher GDPs or countries outside of less developed regions. For all countries in the analysis, we compared their year 2000 GDP per-capita (Penn World Tables 6.2, purchasing power parity, Laspeyres) to two thresholds I$7,000 and I$5,000. In a second set of regressions, we also excluded countries from Latin America or from Eastern Europe and the Former Soviet Union. We found that in wealthier countries in our sample (>I$7,000), the impact of MCV coverage (especially above 80%) attenuated slightly compared to the base case analysis (Table S4A). By contrasting the MCV coefficients from the >I$7,000 and >I$5,000 regressions with the base case, we see that the impact of vaccination was strongest in lower income countries, especially in the range between I$5,000 and I$7,000. For ease of interpretation, we computed the % reduction compared to country-years with no MCV coverage implied by the coefficients at each MCV coverage levels as 100 * [1-exp(β)], where β is the regression coefficient for a particular coverage indicator. For example, the model estimated on countries with per-capita GDPs of >$7,000 and >$5,000 predict reductions in measles deaths of at MCV coverage of 85-94% of 91% and 93% respectively, compared to 92% with the base case specification.
It also appears that in Latin American and Eastern European and Former Soviet countries the impact of MCV was strongest (Table S4B), although differences were substantively negligible. Computing percentage reductions in mortality under different coverage levels as above, we found that the model estimated on countries excluding those in Latin America or excluding Eastern European Former Soviet countries each predicted reductions in measles deaths of 92% at MCV coverage of 90-94%, equivalent to the predicted reduction in the base case specification. In general, excluding poorer countries or those from less developed regions had relatively modest effects on our results. This analysis suggests that our base-case analysis may actually underestimate slightly the potential benefit of increasing MCV coverage in other parts of the world.

Supporting Information S5. Sensitivity analysis: year linear trends versus year fixed effects
The main specification of the model includes calendar year as a linear trend. We find a significant linear trend in year with a slope of -0.117 (Reduction of 11.7% in death rate each year after 1960). As an alternative, we instead used year fixed effects (one dummy variable for each year). In the graph below, the coefficients for each year's fixed effect (and confidence intervals) are plotted. The coefficients can be thought of as the logarithm of the odds ratio and also connote a percent reduction from baseline since the model specification is log-linear. The slope of a line fit through the coefficients (-0.1188) is highly concordant with the year linear trend that was defined prospectively and used in this main base case analysis -suggesting that long-term time patterns are generally captured with the linear trend. Other regression coefficients do not change substantially with the use of the year fixed effect specification (not shown).

Supporting Information S6. Country--years with no observed measles deaths
For analyses of reductions in measles death rates, the rates were log-transformed (natural log). To deal with years with zero observed deaths (for which the log transformation would be undefined), we replaced the 0 with 0.1 times the minimum observed measles-specific death rate in children under 5. As this was a prospectively-defined but arbitrary choice, we explored the effect of alternative replacement values on model results. We replaced country-years having 0 observed deaths with either the minimum observed measles death rate for children under 5 (i.e., a rate 10 times greater than in the main analysis) or else 0.01 times the minimum observed rate (i.e., a rate 10 times smaller than in the main analysis). We then estimated the model using these alternative outcome variables and compared the resulting coefficients for MCV coverage to the coefficients and 95% confidence intervals from the main analysis. Figure S6A shows the results of this sensitivity analysis. The alternative replacements do change the estimated effect, though the magnitude of the change falls within the 95% confidence intervals of the main analysis.
Another alternative would be to use a statistical model for count data such as a Poisson panel regression or a negative binomial panel regression with either fixed or random effects. As Poisson panel regressions are a special case of negative binomial panel regressions in which the mean and variance are assumed to be equal, we estimated negative binomial panel regressions with both fixed and random effects ( Table  S6B). Whereas in the main analysis, MCV coverage of 80% or greater had substantially greater impact than lower levels of coverage, in these alternative models, the effect was more continuous across the coverage levels. Furthermore, at very high coverage levels the reduction compared to country-years with no MCV coverage was estimated to be approximately 80%. The magnitude of the estimated benefit was most consistent with replacing years with no observed deaths with the minimum observed death rates (10 times larger than the main analysis) ( Figure S6A).