We present in this paper a data- and model-driven estimation of the historical measles dynamics, CFR, and vaccination impact in India as a whole as well as in two key states (Bihar and UP) for measles mortality. We used measles mortality data from verbal autopsy studies from a nationally representative longitudinal cohort of 2.4 million households and extracted the periodicity of measles epidemics, which allowed us to infer key parameters driving the dynamics of measles transmission. To do this we used spectral and coherence analysis together with statistical inference using an age-stratified dynamic compartmental model of measles transmission. This represents a first-of-a-kind modeling approach for measles transmission in the developing world.
Spectral analysis of measles mortality data suggested that the time series consisted of the superposition of several cycles of measles epidemics with different inter-epidemic periods. We were able to estimate very high basic reproduction numbers (
R
0 > 20) for the country as a whole and Bihar. The case of UP with its lower estimated basic reproduction number (
R
0 = 14) nevertheless points to the great complexity and potentially significant heterogeneity of measles transmission and mixing contact patterns within the state. In addition, our modeling approach allowed us to estimate a measles CFR between 0.40% and 1.80% in India, which offers a defensible range of CFRs compared with those previously reported in the literature [
7]. It also enabled us to quantify the burden of measles mortality in India, for which we derived estimates in the range of other studies [
24,
25]. Lastly, our analysis confirmed the likely high impact [
26‐
28] of measles SIAs in India which were rolled out for about 110 million Indian children aged 9 months to 10 years [
20] and could have potentially averted 66,000 deaths [
19]. This is also consistent with the significant child survival impact of SIA recently estimated for 25 sub-Saharan African countries [
29].
Our modeling approach gives policymakers tools to estimate the impact of measles routine immunization and SIAs in India, Bihar, and UP. India is highly diverse in terms of geography, health systems, and local epidemiology. Due to a lack of robust mortality data at state and district levels, this analysis modeled an increase in overall vaccine coverage across India to the average level seen in the recent SIA, rather than the details of the SIA itself. As additional data become available, our model will be able to capture finer scale disease dynamics and more accurately assess intervention impact.
Our study presents a number of strengths and limitations. Our model is one of the few dynamic models [
6,
30] of measles transmission calibrated to measles data from low- and middle-income countries; it is to the best of our knowledge the only such model calibrated to measles mortality data taking advantage of the spectral features (periodicity) of measles infection [
31]. As such, it enables us to directly estimate the measles CFR for India, setting itself free from the vast uncertainty in measles CFRs pooled from previous studies [
7]. In general, most attempts at estimating measles mortality and vaccine impact in low- and middle-income countries, and especially in India, have focused on assuming a measles CFR taken from the literature [
5,
6,
27,
28,
32] and on good reporting of measles case notifications during outbreaks, which is unlikely. Hence, our paper proposes a new approach that could strengthen measles mortality estimates and also be applied to other settings where historical time series of measles cases or deaths are available. It also builds on a rich mortality dataset [
8] rather than on under-reported case notifications [
33]. Yet, it has a number of limitations. First, the time series used was short, with only 3 years of measles deaths data, which increased the uncertainty of the spectral estimation. Hence, the longest cycles in the spectrum estimated (with period 1.5 years) may actually be even longer (e.g., 2 years). This is supported by the fact that most measles epidemics observed in the real world have integer-valued periods. Other datasets may present a longer time period; however, they often have a much coarser nature (e.g., WHO measles case notification data [
33]), such as annual data rather than weekly data as with the MDS. This can thus be equally restrictive in the periods calculable with Fourier analysis (Additional file
1: Web appendix III, section 4). Second, it is possible that some measles deaths were not recognized as such in the MDS study, and as a result our CFR estimation would correspond to a lower bound because of such under-ascertainment. However, the MDS dataset has been well validated and scrutinized to reduce biases and misclassifications in the assessment of under-five deaths [
9,
10], so we would anticipate this under-ascertainment to be minimal. Third, small number issues prevented us from examining additional age groups (above age 5) and additional states besides the highly populated states of Bihar and UP. Fourth, our mortality data dated back from 2000 through 2003, which suggests that our extrapolation into the future should be interpreted with caution. For example, the CFR could well decrease over time as under-five mortality decreases and health services improve. Fifth, our modeling assumption of equilibrium behavior is a simplification, because when vaccine coverage and birth rates are changing, transitions in measles epidemic cycles can also occur [
31,
34]. However, this assumption may not be highly inaccurate, because measles coverage in India has been relatively stable in 2000–2003, and the crude birth rate has been only gradually decreasing from 1990 to 2005 [
35,
36].