Table 5.
Comparing the discriminatory accuracy of different dimensions within the health system environment using the proportional change in variance, Zambia 2013–14 (binomial logistic random intercepts model)
| Facility delivery | (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) |
| Reference model | Afford | Cogn | Psych | Geog | Avail | Qual | Afford+ cogn+ psych | Geog+ avail+ qual | |
| ICC | 25% | 23% | 22% | 25% | 14% | 13% | 15% | 20% | 8% |
| PCV | Reference model | −12% | −15% | −4% | −52% | −54% | −47% | −27% | −74% |
| Variance: HS environments | 1.6 | 1.4 | 1.3 | 1.5 | 0.7 | 0.7 | 0.8 | 1.1 | 0.4 |
| (0.6 to 2.8) | (0.5 to 2.6) | (0.5 to 2.5) | (0.5 to 2.8) | (0.2 to 1.4) | (0.2 to 1.4) | (0.2 to 1.6) | (0.3 to 2.2) | (0.1 to 0.9) | |
| Variance: DHS clusters | 1.3 | 1.3 | 1.3 | 1.3 | 1.3 | 1.3 | 1.4 | 1.3 | 1.3 |
| (0.8 to 1.8) | (0.9 to 1.8) | (0.8 to 1.8) | (0.9 to 1.8) | (0.9 to 1.8) | (0.9 to 1.8) | (0.9 to 1.9) | (0.8 to 1.8) | (0.9 to 1.8) | |
| Additive effects (logit coeffs) |
|||||||||
| Afford | −0.8 | −0.9 | |||||||
| (−1.9 to 0.2) | (−1.8 to 0.1) | ||||||||
| Cogn | −1.2 | −1.1 | |||||||
| (−2.3 to −0.1) | (−2.2 to 0.1) | ||||||||
| Psych | −0.9 | −0.3 | |||||||
| (−2 to 0.3) | (−1.5 to 0.8) | ||||||||
| Geog | −2.0 | −1.2 | |||||||
| (−3 to −1) | (−2.2 to −0.2) | ||||||||
| Avail | −1.7 | −0.6 | |||||||
| (−2.6 to −0.9) | (−2.2 to −0.2) | ||||||||
| Qual | −1.8 | −1.0 | |||||||
| (−2.8 to −0.8) | (−1.9 to −0.1) | ||||||||
| Constant | −8.7 | 9.2 | 0.5 | −1.6 | 1.4 | 4.2 | 11.4 | −0.9 | 4.2 |
| (−17.3 to 1.5) | (−6.4 to 21.8) | (−8.6 to 10.8) | (−9.9 to 9.4) | (−6.3 to 9.8) | (−8.2 to 20.1) | (−1.4 to 23.4) | (−13.8 to 7.8) | (−5.9 to 14.1) |
Including a barrier variable in the non-random part of the model in addition to the random part ensures that the HS environments REs’ variance no longer accounts for the additive effect of that variable. This analysis shows the extent to which the ICC decreases with the inclusion of each dimension. A greater decrease in the ICC (and a correspondingly large PCV) indicates that a specific barrier contributes more strongly to the HS environments’ collective discriminatory accuracy. 95% Bayesian credible intervals in parentheses. Controls included in this analysis: mothers’ age at birth, married, secondary school or higher, cluster slope, month-year fixed effects. The model also includes a cross-classified random intercept for DHS sampling clusters in addition to the environments’ random intercept. Individual-level variance is set at 3.29.
Afford, affordability; Avail, availability; Cogn, cognitive; DHS, Demographic Health Survey; Geog, geographic; HS, health system; ICC, intraclass correlation coefficient; PCV, Proportional Change in Variance; Psych, psychosocial; Qual, quality; RE, random effects.