-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathMalaria.Rmd
287 lines (200 loc) · 24.3 KB
/
Malaria.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
---
title: "R & Python Final Project"
author: "Chong Kim"
date: "10/29/2017"
output:
html_document: null
csl: science.csl
resource_files:
- data/final.rds
- data/Moz_admin2.dbf
- data/Moz_admin2.shp
- data/Moz_admin2.shx
- IntProt/server.R
- IntProt/ui.R
- malincpan/server.R
- malincpan/ui.R
- weatherMal/server.R
- weatherMal/ui.R
runtime: shiny
bibliography: bibliography.bib
---
```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE, warning = FALSE, error = FALSE)
library(lattice) # for graphics
library(RColorBrewer)
library(sp)
library(maptools)
library(latticeExtra)
library(rgdal)
library(tmap)
library(cowplot)
library(ggplot2)
library(ggpubr)
library(dplyr)
library(EpiWeek) # for creating epiweek variables
library(tidyverse) # for additional data transformation
library(haven) # for loading data
library(stringr) # for manipulating
library(readr) # for loading data
library(pander) # for reporting table coefficients
library(coda) # gelman
load('analysis.RData')
```
## Introduction
### Burden of Malaria in Africa and WHO efforts
Considerable effort has been made in the past 60 years to identify the precise number of *Plasmodium falciparum* malaria deaths and clinical episodes in Africa [@snow2014sixty]. The major difficulty in identifying the burden has been attributable due to the limitations in health reporting systems. Prior to 2000, the infrastructure in the health systems across Africa was not well established to systematically collect, analyze, and report on the clinical and economic impact of malaria. But all of this changed after setting 'reduction of malaria incidence' as *Target 6c* within the **Millenium Development Goals** (MDG) set by the member states of the United Nations to be achived by 2015. The MDG report of 2015 indicates that throughout the years from 2000 and 2015, over 6.2 million malaria deaths have been averted, global incidence and mortality rate has falled by an estimated 37% and 58% respectively, and more than 900 million insecticide-treated mosquito nets were delivered to malaria-endemic countries in sub-Saharan Africa between 2004 and 2014[@bhatt2015effect].
Of the efforts made by WHO to prevent malaria incidence, insecticide-treated nets (ITN) and indoor residual spraying (IRS) were the main interventions that were distributed across Africa as cheap and effective measures for fighting off malaria. As a short description, ITNs provide further protection compared to non-insecticide-treated nets as they 1) kill mosquitoes, 2) repel mosquitoes from entering the house and 3) reduce the number and length of life of mosquitoes given that high community coverage is achieved. Similarly, IRS provides protection by 1) killing mosquitoes and 2) preventing transmission of infection to other person. To be effective, IRS must be applied to very high proportion of households in an area (usually > 80%).
### Translating WHO effort into Research
From the efforts of the global community, numerous studies[@west2014indoor, @pinder2015efficacy, @lines2015malaria, @deressa2016combining, @ngufor2014combining] have been conducted to determine the impact of preventive interventions (i.e. ITN and IRS) on the incidence of malaria cases in various areas of Africa. Based on the results indicated from the previous studies, the effect of IRS and ITN seemed to be highly dependent on 1) adherence levels of utilization of IRS and ITN, 2) differences in transmission rate among areas and 3) age distribution of individuals. Additionally, different areas of Africa resulted in different results regarding the impact of IRS and ITN on malaria incidence. Specifically, a study conducted in Côte d’Ivoire did not indicate any significant difference between those huts who had ITN vs. those who didn't[@ngufor2014combining]. In comparison, ITN was singnificantly effective in a study conducted in Gambian communities[@pinder2015efficacy].
## Rationale and Objective
Although there has been many studies, as described in the previous section, conducted to confirm and determine the impact of ITN and IRS on malaria incidence, the results do not clearly confirm whether or not there is a statistically significant impact. Additionally, there were no studies that incorporated decaying effects of both ITN and IRS[@wanjala2015insecticidal] in addition to the effects of spatial variables such as community effects[@hawley2003community] or the lagged impact of weather (i.e. @tian2008one and @craig1999climate) on malaria incidence. Specific to the area of our interest, Mozambique, there have not been studies that looked at the incorporating both 1) the decaying effect of ITN and IRS protection such as that of Wanjala et al.[@wanjala2015insecticidal] and 2) the lagged effects of weather on malaria incidence such as that of Ikeda et al.[@ikeda2017seasonally].
In order to extend the research and analyses described above, our goal for this study was to incorporate weather information, decaying impact of ITN and IRS, and other spatio-temporal variables necessary to determine the impact of ITN and IRS on malaria incidence in Mozambique from 2010 to 2017. Specifically this paper investigated 1) the impact of ITN and IRS coverage on malaria incdence over time and 2) whether weather variables had an impact on malaria incidence over time.
## Material and Methods
### Data
#### Incidence Data
Malaria case data for Mozambique, South Africa were collected from January 2010 to July 2017. The incident cases were reported through the local health systems and were collected at the district level at a weekly frequency. Other district level information included district population data, which were obtained from WorldPop, and the 2007 Mozambique census data were used to estimate the proportion of the population $\lt 5$ years of age. Additional geographical information (i.e. squared kilometer and latitude and longitudinal information of district) was also included. The estimated population under age 5 within the district is used as the population at risk during the time period for the poisson regression analysis. Incidence rates were thus calculated by the aggregate incidence within the district, which was multiplied by 1000 to obtain incidence rate per 1000 person-week.
#### Weather Data
Weather data were obtained from the GLobal Land Data Assimilation System and included daily rainfall, minimum, maximum and average temperature, relative humidity, saturation vapor pressure deficit, and surface barometric pressure (a general indicator of large-scale weather activity and exhibits a strong seasonal cycle). Since information was captured daily, average values of the weather data per week were utilized for the poisson regression.
#### Intervention Data
Intervention data was also captured through the local health systems via records based on the MDG efforts mentioned earlier[@plucinski2014evaluation, @lee2017economic]. Intervention data incorporated both ITN and IRS start dates for the districts that were distributed either ITN and/or IRS.
### Analysis
#### ITN and IRS decay
For investigating the association between ITN and IRS protection (scaled from 0 to 1 with 1 being 100% protection) on Malaria incidence, we utilized a decaying rate of protection over time for ITN and IRS with an exponential decay model for both interventions (See Appendix A.1 for details).
#### Lagged Weather
The association between weather data (listed above) and the weekly incidence of malaria was modeled by incorporating 0, 2, 4 and 8 weeks lagged effect of weather on malaria incidence in the poisson regression model.
### Data Analysis
#### Generalized Linear Mixed Model Framework
The estimation of the effect of ITN and IRS protection and weather variables' effect on malaria incidence has been modeled using 1) a standard maximum likelihood estimation approach (via Lapalce Approximation) and 2) a Bayesian approach through MCMC (Markov Chain Monte Carlo) simulation of the posterior distribution. The standard regression model is as such:
\begin{aligned}
g(\mu) & = \log(\Bbb{E}[Y]) = \Bbb{X}\beta + \Bbb{Zb} \\
\end{aligned}
Where the $g(\mu)$ models the mean of the poisson distribution via the natural log link. The covariates that are included in the final model were ITN and IRS protection (scaled from 0 to 1), average temperature (in C$^\circ$), rain (mm/week), Relative Humidity (%), saturation vapor pressure deficit (mmHg), surface barometric pressure (hPa), under 5 total population within the district (as offset variable), time in years, and two interaction terms between time and IRS and time and ITN. The conitnuous variables have been centered and scaled for faster convergence. The weather variables were weekly mean values. The time variable's was specified as a natural cubic spline with $df = 3$.
For the bayeisan estimation of the effects, specification of the location effects ($\beta$ and $u$) are as such:
\begin{aligned}
\begin{bmatrix}
\boldsymbol{\beta} \\
\boldsymbol{u} \\
\boldsymbol{e}
\end{bmatrix} \sim N \left
( \begin{bmatrix}
\boldsymbol{\beta_0} \\
\mathbf{0} \\
\boldsymbol{0}
\end{bmatrix} ,
\begin{bmatrix}
\Bbb{B} & 0 & 0 \\
0 & \Bbb{G} & 0 \\
0 & 0 & \Bbb{R}
\end{bmatrix}
\right )
\end{aligned}
as indicated in [@hadfield2012mcmcglmm]. The $\beta_0$ are the prior means for the fixed effects which were specified as a multivariate normal prior distribution with mean 0 and diagnoal matrix with large variances (1e+10) except for that of the log offset of the under 5 total(i.e. u5total) specified to act as the offset (i.e. $\beta = 1$ and $V = 1e-09$) for calculation of the incidence ratio. The $\Bbb{G}$ and $\Bbb{R}$ structured are specified as an inverse-Wishart distribution with individual elements for each of the components specified as weekly informative priors (The scale matrix of the inverse-Wishart is set to the univariate special case (i.e. inverse-gamma distribution) with a $df = 1$ such that we have a weak prior. see Appendix A.4.2). Same with the maximum likelihood approach, our distribution of the residual and link function will be "poisson" and "log". We have specified 4 chains to be run (in parallel) for mixing of the chains (Appendix A.3.4). For each of the chains we have 3000 burn-in sets with a thinning interval of 10 and sample size per chain set as 1000.
## Result
### Average ITN and IRS Protection Over Time
```{r ITN_IRS, echo=FALSE}
shinyAppDir(
"./IntProt",
options = list(
width = "100%", height = 660
)
)
```
<h1 class="small">**Figure 1.** The figure indicates the average protective effect (scaled from 0 to 1, where 1 is 100% protection) of Insecticide Treated Bed Net (ITN) and Indoor Residual Spraying (IRS) protection throughout the years selected. **LEFT :** Average ITN protection percentage within the year selected. **RIGHT :** Average IRS protection percentage within the year selected.</h1>
Based on the interactive figure that indicates the level of ITN and IRS protection, IRS has a wider coverage, in terms of the intervention being delivered at a greater scale, compared to the ITN's across Mozambique. Areas that had greater protection in the past 3 years (i.e. 2015 - 2017) with regards to ITN are Northern regions such as MALEMA, MECUBURI, NAMAPA_ERATI, MECONTA< and recently MOGINCUAL, whereas from 2013 to 2015 there were higher coverage in the southern and central regions, which does exhibit a greater rate of malaria incidence (Appendix A2.3). Based on initial descriptive analysis, there were 6, 16, 15 and 1 IRS interventions that were delivered to the central, coastal, northern, and souther region respectively. With respect to ITN intervention delivery, there were 34, 56, 49, and 15 interventions were delivered to the central, coastal, northern, and souther region respectively.
### Geographic Distribution of Malaria Incidence Over Time
<center>  </center>
<h1 class="small">**Figure 2.** The figure indicates the Average malaria incidence per 1000 from 2010 to 2017. There isn't much variability over time. The Average incidence rates in each year were below 170 per 1000 people over the range of 2010 to 2017. </h1>
There were 6 districts that had $\gt 100$ cases per 1000 population each year (from pink to yellow color coded districts in Figure 2) from 2010 to 2017. Namely, Chibabava, Cidade de Inhambane, Lalaua, Maravia, Mecufi, and Panda were the districts that had an average of 143, 127, 123, 166, 130, and 147 cases of malaria (per 1000 population) from years 2010 to 2017. There were no distinct variability in the incidence rate throughout the years.
### Impact of Weather on Malaria Incidence per 1000
```{r weath_mal, echo=FALSE}
shinyAppDir(
"./weatherMal",
options = list(
width = "100%", height = 520
)
)
```
<h1 class="small">**Figure 3.** The impact of weather variables on the incidence of malaria. Through the drop-down options, one can choose the impact of 1. Daily rainfall (mm), 2. Daily average temperature (C), 3. Relative humidity (%), 4. Saturdation vapor pressure deficit (mmHg), and 5. Surface barometric pressure (hPa) on malaria incidence. The figure also depicts the 'lagged' effect of each variables. Specifically, The impact of each weather variables with regards to 2 week, 4 week, and 8 week lagged (and No lag) effect on malaria incidence is plotted. </h1>
#### Average Daily Temperature
Based on the smoothed curve (via generalized additive model) displayed in Figure 3, as the average daily temperature (C$^\circ$/week) increases, there was an increasing trend up to 27 C$^\circ$ and then the incidence declines (for all lagged and non-lagged temperature variables) somewhat like a quadratic relationship is displayed. The confidence intervals indicate that the estimated effect of daily average temperature from 20 (C$^\circ$/week) to 28 (C$^\circ$/week) is relatively consistent compared to that of the effect at the lower and upper range of temperature values (i.e. 15 (C$^\circ$/week) < and > 28 (C$^\circ$/week)).
#### Total weekly rainfall
The total weekly rainfall (mm/per week) on the other hand seemed to display a positive linear relationship with malaria incidence. The incidence of malaria was greatest when the (lagged and non-lagged) rainfall amount was at the highest. The 95% confidence interval indicates that the estimated effect of total weekly rainfall on malaria incidence has low variability in the lower range (< 400 mm) but exhibits higher variaiblity towards larger values of rainfall ( > 400 mm).
#### Average relative humidity
Accounting for no lag, 2 weeks lag and 4 weeks lag effect on malaria incidence, as the average relative humidity (%/week) increased from 20% to 90%, the incidence of malaria seemed to increase up to approximately 40 cases/1000 populaiton. However, the relationship between malaria incidence reversed from 90% onwards, as the relative humidity increases a sharp decline in malaria incidence is exhibited. Additionally, when accounting for a lag of 8 weeks the trend seemed to be increasing from 20% to 40% then the incidence of malaria seemed to drop with greater percentage of relative humidity steadily up to 90% then a steep drop in malaria incidence is exhibited similar to the effect of shorter lag periods described above. The 95% confidence intervals indicate that the impact of average relative humidity on malaria incidence has greater variability at lower values of humidity (< 40%).
#### Average saturation vapor pressure deficit
As average saturation vapor pressure (mmHg/week) increases up to 17 mmHg, there is a steady increase in malaria incidence (especially accounting for a 8 week lag), then as saturation vapor pressure increases up to 30 mmHg malaria incidence seems to decreases steadily. The concave down relationship is displayed for all saturation vapor effect (i.e. no, 2 weeks, 4 weeks, and 8 weeks lag) on malaria incidence. Similar to the effect of average weekly rainfall, the 95% CI of the effect of saturation vapor pressure deficit (mmHg) on malaria incidence has small variability at lower values ( < 20mmHg) but has greater variability at larger values ( > 20 mmHg).
#### Average surface barometric pressure
The average surface barometric pressure (hPa/week) displays a highly non-linear relationship with malaria incidence. There is an initial steep increase in malaria as the pressure increases from 870 to 912 then a steep decline afterwards. Overall a fluctuating (i.e. sinusoidal) relationship is displayed. There is very little variability in in terms of the effect of surface barometric pressure on malaria incidence except at the upper limite (1015 hPa < ).
### Impact of Intervention on Malaria Incidence per 1000
```{r int_inc_int, echo=FALSE}
shinyAppDir(
"./malincpan",
options = list(
width = "100%", height = 430
)
)
```
<h1 class="small">**Figure 4.** The figure indicates the impact of Insecticide Treated Bed Net (ITN) and Indoor Residual Spraying (IRS) protection on malaria incidence (scaled from 0 to 1, where 1 is 100% protection). **TEAL :** The impact of ITN protection on malaria incidence. **RED :** The impact of IRS protection on malaria incidence.</h1>
With regards to the effect of ITN protection, there is a concave upwards dip in the lower protection values (i.e. from 0 to 0.3) then afterwards the overall relationsihp between ITN protection and malaria incidence is constant (at 35 cases per 1000 population).
Similarly, IRS protection displays a much larger dip and then a steep incline from values 0 to 0.11 (11% IRS protection) then a fluctuating relationship until 75% protection value then as protection increases the incidence decreases steadily (from 0.75 to 1).
### Generalized Linear Mixed Effect Model Regression Analysis
#### Table 1. GLMM Parameter Estimates with 0, 2, 4, and 8 weeks lagged effect of Weather
```{r mcmcresult, echo=FALSE}
panderOptions('table.split.table', Inf)
resultab<- cbind(summary(models[[1]])$statistics[1:12,1:2], summary(models[[2]])$statistics[1:12,1:2], summary(models[[3]])$statistics[1:12,1:2], summary(models[[4]])$statistics[1:12,1:2])
rownames(resultab) <- c("Intercept", "ITN", "IRS", "Avg. Temp (C)", "Avg Rain (mm/week)", "Rel. Humidity (%)", "Vapor Pressure (mmHg)","Surface Pressure (hPa)", "Under 5 Total", "Time", "ITN:Time", "IRS:Time")
colnames(resultab) <- c("M1 mean", "M1 SD", "M2 mean", "M2 SD", "M3 mean", "M3 SD", "M4 mean", "M4 SD")
pander(resultab)
```
<h1 class="small">\* M1: No lagged effect of weather variables; M2: 2 week lagged effect of weather variables; M3: 4 week lagged effect of weather variables; M4: 8 week lagged effect of weather variables.</h1>
#### Table 2. GLMM Variance Component Estimates
```{r varcomp, echo=FALSE}
panderOptions('table.split.table', Inf)
pander(retab)
```
```{r coefplot, echo = FALSE, fig.align='center', fig.width=7, fig.asp=0.8}
# plot all 4 MCMC model coefficients
plot_mcfix <- function(y) {
# extract each of the MCMC models from list and get parameter estimates
x <- summary(y[[1]])
x2 <- summary(y[[2]])
x3 <- summary(y[[3]])
x4 <- summary(y[[4]])
# creating plot scheme
#n <- dim(x$statistics)[1]
n <- 12
par(mar=c(4, 10, 4, 1))
plot(x$statistics[1:12,1], n:1,
yaxt="n", ylab="", xlab = "Parameter estimates",
xlim=c(-5,3),
pch=19,
main="Posterior means and 95% credible intervals", col=c("red"))
points(x2$statistics[1:12,1], n:1, pch=19, col = "blue")
points(x3$statistics[1:12,1], n:1, pch=19, col = "green")
points(x4$statistics[1:12,1], n:1, pch=19, col = "yellow")
grid()
axis(2, at=n:1,
rownames(x$statistics)[1:12] <- c("Intercept", "ITN", "IRS", "Avg. Temp (C)", "Avg Rain (mm/week)", "Rel. Humidity (%)", "Vapor Pressure (mmHg)","Surface Pressure (hPa)", "Under 5 Total", "Time", "ITN:Time", "IRS:Time")
,
las=2) # originally rownames(x$statistics)
arrows(x$quantiles[1:12,1], n:1, x$quantiles[1:12,5], n:1, code=0)
abline(v=0, lty=2)
legend("bottomleft", legend = c("Model 1", "Model 2", "Model 3", "Model 4"), lty = c(1,2,3,4), col = c("red","blue","green","yellow"))
}
par(mfrow=c(1,1))
plot_mcfix(models)
```
<h1 class="small">**Figure 4.** The figure indicates the impact of Insecticide Treated Bed Net (ITN) and Indoor Residual Spraying (IRS) protection on malaria incidence. **TEAL :** The impact of ITN protection on malaria incidence. **RED :** The impact of IRS protection on malaria incidence.</h1>
Accounting for the between-district variability (as random effects), there wasn't a big difference in the effect of weather and interventions across the 4 models that also accounted for different levels of lagged effects of weather variables. Based on the model that incorporates the effect of 8 weeks lagged weather variables impact on malaria incidence (Model 4), there was a statistically significant effect of average daily temperature, average relative humidity, and saturation vapor pressure deficit on the incidence of malaria per 1000. Specifically, on average effect of a unit increase in the average daily temperature (in C$^\circ$) is `r round(exp(summary(models[[4]])$statistics[,1][['scale(tavg_lag_8)']]),2)` times greater rate of malaria incidence per 1000 whereas a unit increase in the average relative humidity (in %/week), average saturation vapor pressure deficit (mmHg/week) is `r round(exp(summary(models[[4]])$statistics[,1][['scale(rh_lag_8)']]),2)` and `r round(exp(summary(models[[4]])$statistics[,1][['scale(sd_lag_8)']]),2)` times smaller rate of malaria incidence per 1000, respectively. The rate of malaria incidence did not have a statistically significant association with ITN and IRS over time as indicated in FIgure 4.
The between-district variabilility (Table 2) was 20-folds greater than the residual variance indicating that there was significant district level differences in malaria incidence which was also supported through prediction intervals and profiled zeta pair plots (in Appendix).
## Conclusion & Discussion
We have shown that there are lagged effects of weather factors in addition to significant between-district variability on the inicidence of malaria in Mozambique. Results from the generalized linear mixed model analysis have shown that there were significant associations between malaria incidence and weather variables, however there were no statistically significant association between ITN and IRS interventions with the incidences.
Our results here support previous studies that linked malaria incidence to climate and weather phenomena, but also indicates that there potentially is no effect of ITN and IRS interevention effects on malaria incidence. However, this study should not be interpreted as an ultimatum on the effect of of the weather and intervention on the incidence of malaria due to the lack of other possible modeling approaches that have been taken in this analysis. Specifically there are other potential factors (confounders) that may affect both exposure and outcome that are were not included in the analysis such as urbanicity of the districts, vegetation type and grwoth stage[@diggle2007spatial], health service infrastucture of district, and many more. Another limitation to the study is that the approach for analysis utilized a relatively simple model that incorporated only one random effects term (district) with one interaction effect (time and intervention) whereas other effects such as region could be incorporated as another heirarchical cluster. Our approach also did not allow for potential non-linear effects of weather and intervention on malaria incidence, which further analyses should explore.
With regards to future analyses, the three way interaction between weather, time, and intervention should be explored, in addition to incorporating random slopes and intercepts for the three variables. Additionally, other potential models that should be explored further are self-organizing maps (SOMs) [@ikeda2017seasonally], gaussian point process [@diggle2007spatial, @diggle2013statistical], and machine learning methods (e.g. Kernal Principal COmponent Analysis and Support Vector Machines [@gokaraju2011machine]).
## Acknowledgement
The authors thank Dr. Kathryn Colborn and Dr. Debashis Ghosh for their amazing teaching in the Python & R Programming in Data Science Fall 2017 (BIOS 6640) course. This work was not technically supported by any organization but practically the Department of Pharmaceutical Sciences (DOPS) and Department of Biostatistica and Bioinformatics (DOBB) has provided numerous educational supplies and courses that have allowed for this project to be completed as such.
## References
<div id="refs"></div>
## Appendix
Appendix file can be found in https://github.com/ck2136/malaria2k17.