The provider assessment captured data on readiness and quality of care 144 public first-level facilities, 43 religious or private first-level facilities, and seven referral facilities (5 public and 2 private).
The MICS captured data on care-seeking for sick children and labor and delivery care from 44 household clusters in the Savanes region. This sample included 392 women who gave birth in the two years preceding the survey, among whom 65.2% (95% CI: 53.2-75.5%) delivered at a health facility. The sample also included 183 children under five who experienced fever, diarrhea, or symptoms of ARI in the two weeks preceding the survey. Among these children, care was sought for 43.2% (35.7-60.0%).
Fig 1. Source of sick child and delivery care reported in household survey
The most common source of care for sick children (Fig 1) were public first-level facilities (27.0% of sick children), followed by pharmacies (8.2%). In our analysis, we considered pharmacies unskilled providers, we treated these children as receiving no skilled care if they did not seek care from another source. Just under half (46.0%) of women reported delivering at a first-level public facility (including maternity wards), followed by public referral facilities and private first-level facilities. No women reported seeking care from private referral facilities for delivery or sick child care. Using the exact-match linking method, 93.9% of sick children and 92.3% of delivering women could be assigned to their stated source of care.
Real Facility Quality Data
Figure 2 shows the distribution of facility inputs and quality scores by provider type for sick child and delivery care. On average, referral facilities offer greater structural and process quality than first-level facilities. There is variation in individual first-level providers' quality, although the interquartile range (25th to 75th percentile) was typically less than 20 percentage points.
Fig 2. Distribution of facility scores, original Cote d'Ivoire data
In figures 3 and 4, we compare the effective coverage estimate for labor and delivery care and sick child care, respectively, based on a census of providers using two linking methods (Euclidean distance or administrative unit average score) against the estimate generated using each random sample (n=20) by linking method and sample size. Except for input-adjusted coverage for sick child care, the administrative linking method slightly overestimated exact-match coverage, and the Euclidean distance linking method slightly underestimated exact-match coverage. However, both ecological linking methods using a census of providers produced estimates of input- and quality-adjusted coverage that did not vary significantly from the exact-match linking estimates.
Fig 3. Estimates of input- and quality-adjusted coverage of labor and delivery care using original quality scores, by ecological linking method and facility sample size
<The red line indicates the input- or quality-adjusted coverage based on exact match linking. Light red lines indicate the 95% CI around the exact-match coverage estimate. Green (input-adjusted) and blue (quality-adjusted) dots indicate the coverage estimate derived from the ecological linking method applied to the census of providers. Light green (input-adjusted) and light blue (quality-adjusted) bars indicate the 95% CI around the census-derived ecologically linked effective coverage estimates. Gray dots indicate the effective coverage estimates produced using each of the 20 samples by sample size and ecological linking method.>
Fig 4. Estimates of input- and quality-adjusted coverage of sick child care using original quality scores, by ecological linking method and facility sample size
All sample-derived estimates fell within the confidence bounds of both the census-based ecological linked estimates and the exact-match estimates for labor and delivery care (Fig 3). A greater spread in sample-derived estimates were observed for the smaller samples (n=65), however, these estimates were still within the census estimates' bounds. We observed similar results with sick child care effective coverage estimates (Fig 4). Despite greater spread in point estimates with smaller sample sizes, all sampled estimates fell within the confidence bounds of both the ecologically-linked census estimates and the exact-match linked estimates. Table 1 summarizes the effective coverage estimates using each of the linking and sampling methods.
Table 1. Summary of input- and quality-adjusted coverage estimates by linking method, facility sample size, and dataset
|
|
Input-Adjusted Coverage
|
Quality-Adjusted Coverage
|
LABOR AND DELIVERY
|
|
Admin Link
|
Nearest Link
|
Admin Link
|
Nearest Link
|
ORIGINAL
|
|
|
|
|
|
Exact Match
|
Mean (95% CI)
|
37.2 (30.5-43.9)
|
40.1 (32.9-47.3)
|
Census
|
Mean (95% CI)
|
38.8 (31.9-45.7)
|
37.0 (30.0-44.0)
|
40.8 (33.6-48.0)
|
39.6 (32.0-47.1)
|
Sampled Datasets
|
|
|
|
|
|
N=65
|
Median [IQR]
|
38.4 [37.9-39.0]
|
37.1 [36.3-37.7]
|
40.4 [39.9-41.2]
|
40.2 [39.8-40.5]
|
N=90
|
Median [IQR]
|
38.6 [38.3-38.9]
|
37.2 [36.2-37.8]
|
40.5 [40.2-41.0]
|
40.3 [39.6-40.9]
|
N=130
|
Median [IQR]
|
38.8 [38.5-39.0]
|
37 [36.8-37.3]
|
40.6 [40.4-41.0]
|
39.8 [39.6-40.4]
|
RANDOM
|
|
|
|
|
|
Exact Match
|
Mean (95% CI)
|
32.9 (25.5-40.2)
|
35.3 (27.4-43.2)
|
Census
|
Mean (95% CI)
|
36.1 (29.3-42.9)
|
32.8 (24.8-40.8)
|
34.6 (28.2-41.0)
|
32.3 (24.1-40.5)
|
Sampled Datasets
|
|
|
|
|
|
N=65
|
Median [IQR]
|
36.0 [33.8-36.8]
|
35.3 [32.7-37.1]
|
34.8 [33.4-36.0]
|
32.6 [29.7-34.0]
|
N=90
|
Median [IQR]
|
35.8 [34.6-36.8]
|
34.5 [33.0-35.4]
|
34.1 [32.8-35.1]
|
32.2 [30.2-33.5]
|
N=130
|
Median [IQR]
|
35.9 [34.7-36.9]
|
34.5 [33.6-35.1]
|
34.7 [34.0-35.2]
|
32.6 [31.6-33.6]
|
PREFERENTIAL CARE-SEEKING
|
|
|
|
|
|
Exact Match
|
Mean (95% CI)
|
45.1 (37.0-53.2)
|
48.0 (39.4-56.7)
|
Census
|
Mean (95% CI)
|
43.5 (35.6-51.4)
|
43.2 (34.5-51.8)
|
45.2 (37.1-53.4)
|
45.7 (36.6-54.8)
|
Sampled Datasets
|
|
|
|
|
|
N=65
|
Median [IQR]
|
43 [42.1-43.8]
|
39.2 [38.2-40.8]
|
43.6 [42.8-45.8]
|
42.1 [41.1-43.4]
|
N=90
|
Median [IQR]
|
43.1 [42.9-43.8]
|
40.8 [39.5-42.2]
|
44.6 [43.5-45.5]
|
43.2 [42.5-45.4]
|
N=130
|
Median [IQR]
|
43.5 [43-44]
|
40.9 [39.9-42.4]
|
45 [44.5-45.5]
|
44.4 [42.4-45.6]
|
|
|
|
|
SICK CHILD
|
|
|
|
|
|
ORIGINAL
|
|
|
|
|
|
Exact Match
|
Mean (95% CI)
|
20.8 (15.9-25.6)
|
16.8 (12.8-20.8)
|
Census
|
Mean (95% CI)
|
20.0 (15.3-24.8)
|
19.7 (14.8-24.6)
|
17.2 (13.1-21.3)
|
16.4 (12.16-20.7)
|
Sampled Datasets
|
|
|
|
|
|
N=65
|
Median [IQR]
|
20.1 [19.7-20.5]
|
19.7 [19.3-20.3]
|
17.3 [16.9-17.5]
|
17.2 [16.6-17.6]
|
N=90
|
Median [IQR]
|
19.9 [19.8-20.3]
|
19.7 [19.4-20.0]
|
17.2 [17.0-17.6]
|
16.7 [16.5-17.2]
|
N=130
|
Median [IQR]
|
20.2 [19.9-20.3]
|
19.7 [19.5-20.0]
|
17.2 [17.0-17.4]
|
16.7 [16.5-17.1]
|
RANDOM
|
|
|
|
|
|
Exact Match
|
Mean (95% CI)
|
19.8 (14.5-25.1)
|
19.2 (14.7-23.8)
|
Census
|
Mean (95% CI)
|
19.9 (14.7-25.1)
|
19.4 (13.8-25.0)
|
18.8 (13.7-23.9)
|
18.4 (13.0-23.8)
|
Sampled Datasets
|
|
|
|
|
|
N=65
|
Median [IQR]
|
19.9 [19.3-20.7]
|
18.2 [17.3-19.5]
|
18.1 [17.6-19]
|
18 [17-18.4]
|
N=90
|
Median [IQR]
|
19.6 [19.3-20.5]
|
18.7 [17.7-20.2]
|
18 [17.8-18.9]
|
17.8 [17.1-18.7]
|
N=130
|
Median [IQR]
|
20.1 [19.5-20.3]
|
19 [18.6-19.4]
|
18.4 [17.9-19.1]
|
17.8 [17.1-18.3]
|
PREFERENTIAL CARE-SEEKING
|
|
|
|
|
|
Exact Match
|
Mean (95% CI)
|
25.6 (19.6-31.5)
|
21.6 (16.5-26.6)
|
Census
|
Mean (95% CI)
|
21.2 (16.0-26.4)
|
21.9 (16.1-27.7)
|
18.4 (13.9-23.0)
|
18.6 (13.4-23.9)
|
Sampled Datasets
|
|
|
|
|
|
N=65
|
Median [IQR]
|
21.1 [20.6-21.6]
|
20.9 [20.3-21.7]
|
18.6 [17.9-19.4]
|
18.6 [17.7-20.1]
|
N=90
|
Median [IQR]
|
21.1 [20.5-21.3]
|
20.8 [20.3-21.5]
|
18.8 [18.5-19.3]
|
18.5 [17.6-19.2]
|
N=130
|
Median [IQR]
|
21.1 [20.8-21.4]
|
21.4 [20.4-22.1]
|
18.6 [17.9-19]
|
18.6 [18.2-18.9]
|
Simulated Random Facility Quality Data
Assigning each facility a structure and process-quality score at random, we estimated the effect of sampling facilities on effective coverage estimates in settings of high variability in provider quality. Figure 5 shows the distribution of facility input and quality scores for the facility census with random quality assignment. As expected, the median scores across for the two most numerous provider types (public and private first-level facilities) were approximately 50%, with an interquartile range of roughly 25% and 75%, respectively.
Fig 5. Distribution of facility scores, random index scores
In figures 6 and 7, we compare the effective coverage estimate based on a census of providers using the two linking methods against the estimates generated using the samples of random index scores (n=20) by linking method and sample size. All sample-derived estimates fell within the bounds of their respective census-derived ecologically linked value. For labor and delivery, one sampled estimate of input-adjusted coverage and quality-adjusted coverage each fell outside of the confidence bounds of the exact-match coverage estimate. Both outlying estimates were a product of the smallest (n=65) sample size. None of the sampled estimates of input- or quality-adjusted coverage of sick child care fell outside the exact-match estimate bounds. Input-adjusted coverage estimates derived from samples using Euclidean distance appear to over-estimate labor and delivery coverage and underestimate sick child care when comparing the median and IQR of sampled estimates against the census-derived estimates (Table 1). The spread in sampled estimates was generally greater for Euclidean distance-derived estimates compared to administrative average estimates.
Fig 6. Estimates of input- and quality-adjusted coverage of labor and delivery care using random quality scores, by ecological linking method and facility sample size
Fig 7. Estimates of input- and quality-adjusted coverage of sick child care using random quality scores, by ecological linking method and facility sample size
Simulated Preferential Care-seeking from Higher Quality Providers
After simulating preferential care-seeking from higher quality providers using the original Cote d'Ivoire data set, we see a greater spread in structural and process quality scores within provider categories compared to the original dataset. We increased the score of the more heavily utilized facilities, while less used facilities have reduced scores. As in the original dataset, the simulated scores for referral facilities continue to be greater than first-level facilities. However, the interquartile range in first-level providers' scores increased to approximately 25 to 45 percentage points (Fig 8). The median score for public first level facilities, the most utilized source of care, remained stable.
Fig 8. Distribution of facility scores, simulated preferential care-seeking from higher-quality providers
Ecological linking systematically underestimated the exact-match estimates using this preferential care-seeking simulated data set (Fig 9-10). The census and sample based ecological linking underestimated the exact-match point estimate for all linking methods, sample sizes, and outcomes. However, few estimates were statistically different from the exact-match coverage estimates. The smallest sample size resulted in nine estimates falling below the exact-match estimate's lower bound (across health areas and outcomes). Both the moderate (n=90) and large (n=130) sample sizes resulted in two outliers each. More outliers occurred using the Euclidean distance linking method (n=11) versus the administrative linking (n=2). The two administrative linking sampled estimates that fell outside the exact-match estimate bounds were estimates of input-adjusted sick child care estimated using a sample size of 65 and 90. Census estimates generated using the administrative linking approach were similarly biased against the exact-match estimate, compared to the Euclidian distance method. However, the sampled estimates derived using the Euclidian linking approach were more variable as demonstrated by the number outliers generated using the approach and the wider IQR of sampled estimates comparing Euclidean versus administrative linking methods by sample size and outcome (Table 1).
Fig 9. Estimates of input- and quality-adjusted coverage of labor and delivery care using simulated preferential care-seeking scores, by ecological linking method and facility sample size
Fig 10. Estimates of input- and quality-adjusted coverage of sick child care using simulated preferential care-seeking scores, by ecological linking method and facility sample size
While 2.7% of our sampled estimates fell outside of the exact-match bounds, none of the sampled estimates fell outside of the bounds of the census-derived estimates generated through ecological linking. As observed with the other data sets, there was a greater spread in sample-derived estimates generated using smaller facility sample sizes.