Does adjusting for recall in trend analysis affect coverage estimates for maternal and child health indicators? An analysis of DHS and MICS survey data

Background: The Demographic and Health

Surveys (DHS) and Multiple Indicator Cluster Surveys (MICS) are the

major data sources in low- and middle-income countries (LMICs) for

evaluating health service coverage. For certain maternal and child

health (MCH) indicators, the two surveys use different recall

periods: 5 years for DHS and 2 years for MICS. Objective: We

explored whether the different recall periods for DHS and MICS

affect coverage trend analyses as well as missing data and coverage

estimates. Designs: We estimated coverage, using proportions with

95% confidence intervals, for four MCH indicators: intermittent

preventive treatment of malaria in pregnancy, tetanus vaccination,

early breastfeeding and postnatal care. Trends in coverage were

compared using data from 1) standard 5-yearDHS and 2-year MICS

recall periods (unmatched) and 2) DHS restricted to 2-year recall

to match the MICS 2-year recall periods (matched). Linear

regression was used to explore the relationship between length of

recall, missing data and coverage estimates. Results: Differences

in coverage trends were observed between matched and unmatched data

in 7 of 18 (39%) comparisons performed. The differences were in the

direction of the trend over time, the slope of the coverage change

or the significance levels. Consistent trends were seen in 11 of

the 18 (61%) comparisons. Proportion of missing data was inversely

associated with coverage estimates in both short (2 years) and

longer (5 years) recall of the DHS (r0.3, p0.02 and r0.4, p0.004,

respectively). The amount of missing information was increased for

longer recall compared with shorter recall for all indicators

(significant odds ratios ranging between 1.44 and 7.43).

Conclusions: In a context where most LMICs are dependent on

population-based household surveys to derive coverage estimates,

users of these types of data need to ensure that variability in

recall periods and the proportion of missing data across data

sources are appropriately accounted for when trend analyses are

conducted