Table 3. Impact of PMI and sensitivity analysis on primary and secondary outcome measures of interest using different matching procedures.
| Matching procedures and sensitivity analysis | U5MR per 1000 live birth | ITN (%) | IRS (%) | ACT (%) |
|---|---|---|---|---|
| ATET (95% CI) | ATET (95% CI) | ATET (95% CI) | ATET (95% CI) | |
| Matching procedures | ||||
| Mahalanobis metric matching on covariates within propensity score caliper: Variables used in estimating Mahalanobis metric for U5MR includes IRS, ITN, and ACT | -10.50 (-18.66, -2.33); p = 0.012 |
|||
| Mahalanobis distance metric with nearest neighbor matching: 1:1: Biases adjusted for population size and country | -11.86 (-20.61, -3.10); p = 0.008 |
0.23 (0.17, 0.30) P<0.001 |
0.08 (0.06,0.11) P<0.001 |
0.11 (0.08, 0.13); P<0.001 |
| Mahalanobis distance metric with nearest neighbor matching: 2:1: Biases adjusted for population size and country | -12.65(-20.48, -4.817); p = 0.002 |
0.23 (0.18, 0.29); p<0.001 |
0.08 (0.06,0.10); p<0.001 |
0.10 (0.08, 0.13) P<0.001 |
| Sensitivity Analysis | ||||
| Coarsened Exact Matching with propensity score: Checking imbalance on corruption | -9.60 (-20.11, 0.89); p = 0.073 |
0.13 (0.08,0.18), p<0.001 | 0.07 (0.02, 0.12); p = 0.005 | 0.07 (0.04,0.10); p<0.001 |
| Coarsened Exact Matching | -13.83 (-21.98, -5.68); p = 0.001 |
0.16 (0.11,0.21); p<0.001 | 0.04 (0.00, 0.08); p = 0.034 | 0.05 (0.02,0.08); p<0.001 |
| Subclassification on propensity score: Five subclasses | -17.34 (-24.45, -6.50); p<0.05 |
0.18 (0.12, 0.21); p<0.05 | 0.04 (-0.01, 0.07); p>0.05 | 0.08 (0.07; 0.10) p<0.05 |
U5MR: Under-five mortality rate; Population coverage of insecticide-treated nets (ITNs): Proportion of people who slept under insecticide-treated bednet on any given night each year; Population coverage of indoor residual spraying (IRS): Proportion of the population protected by indoor spraying of insecticides (IRS); Population coverage of artemisinin-based combination therapy (ACTs): Proportion of fever cases in under-five-year-olds receiving ACTs. Variance in the treatment group is much larger than that in the control group, smaller calipers were necessary as indicated by Rosenbaum and Rubin (1985b). The authors therefore used a caliper of 0.25 standard deviations of the linear propensity score: CI: Confidence interval based on robust standard error, ATET: Average PMI effect on recipient countries. For Mahalanobis with K: 1 Nearest neighbor matching, biased was adjusted for population size and country, PSM: Propensity score matching; CEM: Coarsened Exact Matching: Estimating the sample average treatment on the treated (the SATT), P-value notation: p<0.05 statistically significant.