Skip to main content

Table 6 Performance of predictive generalised linear models in South Sudan, by acute malnutrition outcome

From: Can we predict the burden of acute malnutrition in crisis-affected countries? Findings from Somalia and South Sudan

Statistic Categorical outcomes Continuous outcomes
SAM (WFH + oedema) GAM (WFH + oedema) SAM (MUAC + oedema) GAM (MUAC + oedema) WFH MUAC
Predictors: coefficient† on training data
Incidence of acute flaccid paralysis previous 3mths   previous 3mths previous 3mths   previous 3mths
 0 per 100,000 [ref.]   [ref.] [ref.]   [ref.]
 0.01 to 0.49 per 100,000 0.94   0.80* 0.99   0.08**
 ≥ 0.50 per 100,000 1.17*   1.19 1.16**   -0.03
Main local livelihood type
 Agriculturalists   [ref.]    [ref.] [ref.]
 Agro-pastoralists   1.46***    -0.30*** -0.21***
 Displaced   1.44***    -0.36*** -0.09*
 Pastoralists   1.02    -0.15*** -0.13**
Total rainfall previous 6mths previous 6mths    previous 6mths  
 < 50mm [ref.] [ref.]    [ref.]  
 50 to 99mm 0.98 0.99    -0.02  
 100 to 149mm 0.74*** 0.80***    0.11***  
 ≥ 150mm 0.78 0.97    0.00  
Terms of trade (flour-goat exchange) 3-5mths prior 4-6mths prior    4-6mths prior  
 < 20.0Kg [ref.] [ref.]    [ref.]  
 20.0 to 29.9Kg 0.88* 0.80***    0.15***  
 30.0 to 39.9Kg 0.70*** 0.84***    0.10***  
 ≥ 40.0 Kg 0.81** 1.00    0.02  
Incidence of measles   2-4mths prior   1-3mths prior 1-3mths prior 1-3mths prior
 0   [ref.]   [ref.] [ref.] [ref.]
 > 0   1.06   1.16* -0.04 -0.07*
Doses of measles vaccine administered    3-5mths prior    
 0 per 100,000    [ref.]    
 0.1 to 99.9 per 100,000    1.24    
 100.0 to 199.9 per 100,000    1.21    
 200.0 to 299.9 per 100,000    1.74***    
 ≥ 300.0 per 100,000    1.14    
Estimation performance
Mean square error training data 0.00046 0.00297 0.00017 0.00198 0.05358 0.10680
LOOCV 0.00056 0.00368 0.00020 0.00221 0.06670 0.12667
holdout data 0.00039 0.00342 0.00009 0.00150 0.06680 0.09943
Relative bias LOOCV +41.1% +12.7% +80.9% +38.9% +9.9% 0.0%
holdout data +49.1% +14.4% +88.4% +20.5% +16.0% -0.1%
Relative precision of 95%CI LOOCV ±11.4% ±5.4% ±16.4% ±4.8% ±2.9% ±0.2%
holdout data ±11.4% ±6.1% ±16.1% ±4.6% ±3.3% ±0.2%
Coverage of 95%CI LOOCV 64.3% 45.1% 68.7% 49.6% 38.9% 30.4%
holdout data 66.1% 58.9% 82.1% 57.1% 30.4% 48.2%
Coverage of 80%CI LOOCV 44.3% 29.2% 47.0% 34.8% 28.3% 26.1%
holdout data 46.4% 33.9% 60.7% 42.9% 21.4% 33.9%
Classification performance by SAM/GAM prevalence threshold (n = denominator of percentage)
Sensitivity, lower threshold LOOCV ≥2% 100.0% (97) ≥15% 89.7%
(78)
≥2% 28.1%
(32)
≥15% 0.0%
(9)
n/a
holdout data 100.0%
(48)
93.8%
(32)
27.3%
(11)
0.0%
(4)
Sensitivity, upper threshold LOOCV ≥5% 31.0%
(42)
≥20% 60.0%
(45)
≥5% 0.0%
(3)
≥20% 0.0%
(2)
holdout data 42.9%
(14)
35.3%
(17)
n/a
(0)
0.0%
(2)
Specificity, lower threshold LOOCV <2% 0.0%
(18)
<15% 40.0%
(35)
<2% 86.7%
(83)
<15% 100.0%
(106)
holdout data 0.0%
(8)
25.0%
(24)
100.0%
(45)
100.0%
(52)
Specificity, upper threshold LOOCV <5% 76.7%
(73)
<20% 76.5%
(68)
<5% 100.0%
(112)
<20% 100.0%
(113)
holdout data 90.5%
(42)
87.2%
(39)
100.0%
(56)
100.0%
(54)
  1. †Odds ratio for categorical outcomes; linear coefficient for continuous outcomes
  2. *0.01 ≤ p-value < 0.05 ** 0.001 ≤ p-value < 0.01 *** p-value < 0.001