Regional inequality and vaccine uptake: a multilevel analysis of the 2007 Welfare Monitoring Survey in Malawi

The analysis in this article is based on data from the 2007 Welfare Monitoring Survey (WMS) in Malawi which was one of eight modules in the National Census of Agriculture and Livestock (NACAL) collected by the National Statistical Office (NSO). The survey took place autumn 2007 over a period of about two weeks and data on 21 090 children below 5 years are included in the survey. In this study, we restricted our analysis among those children ages 10–60 months (N=18 251), i.e. the sample population.

Since the main purpose of the NACAL was related to agriculture, the first sampling step included only households with land and animals. For the WMS module, additional landless households were selected so that the probability of being selected for the WMS was the same for households with and without land. The sampling of enumeration areas included both urban and rural areas, further documentation can be fond on the Malawi National Statistical Office webpage [23].

The household head was asked questions about all family members, which makes it possible to decide whether there are other children in the household than the household head’s own children, whether any of the children are orphans and whether the parent or guardian has an education. Questions about the vaccination status of all the children in the sample population of the household were asked. Variables on housing and health conditions, poverty indicators and distance to important infrastructure are also included. Since the child and household information is connected to the NACAL village module, vaccination coverage may also be tracked down to region and village level and to information of family structure and ethnicity.

In addition to the WMS, regional data on delivery attended by health care personnel were added to the dataset. We also included a regional variable reporting the percentage of the population living in a permanent dwelling as a general indicator of regional variations in level of living conditions. The regional data is documented in a publication issued by the National statistical office in Malawi [24].

All households involved in the survey were asked questions for each child less than 5 years on whether it received measles, DPT (three doses), Polio (4 doses) and BCG vaccinations. The response was based both on information from vaccination cards (88%) and/or from mother/guardians recollection (12%). It is specified whether a vaccination card is shown, but no information exists on when the vaccination was given. This makes it impossible to decide whether the child got the vaccination at the recommended age. Without vaccination cards, recall bias might be an issue. We decided to include families without a health card. The decision was informed by previous assessments of the quality of child immunization coverage estimates in population based surveys, that found no major systematic weakness in recall data [25]. Full immunization coverage included children that have received all of the nine vaccinations. The outcome variable is a dichotomy, identifying children with full vaccination coverage and children with less than full coverage.

Characteristics related to child include gender and year of birth. Two variables indentify mother’s school attendance and literacy status (able to read and write). Mother’s marital status consist 6 different categories, never married, married monogamous, married polygamous, divorced, separated and widowed. The study also includes a poverty estimate that divides the households into five quintiles based on the estimated annual household consumption per capita in the WMS. The estimate has been calculated using a poverty model [26] and is independent of the variables used to explain the variation in vaccination coverage in this study.

Health care experiences include two variables related to the birth of the child: where the child was delivered (five categories: hospital/maternity clinic, health clinic, health centre, health post, at home, other), and who assisted with the delivery of the child (four categories: doctor/clinical officer, midwife/nurse, trained traditional birth attendant (t.b.a), self). A variable describing walking distances to the nearest health care facility was also included.

Regional level variables comprise information about the percentage of deliveries assisted by trained personnel, and percentage of the population living in permanent dwellings. The sources of these variables are not from the 2007 WMS, but based on the 2008 Malawi Population and Housing Census, documented in Table 1.

The data were analyzed using two-level logistic regression models (mixed models), where individual-level factors considered as lower-level predictors and the regional-level factors were considered as higher-level predictors. We estimated both fixed and random effects; the first explains the variation of vaccine coverage at the individual-level and the later explains the variation of vaccine coverage at the regional-level. First, we developed random intercept models, explaining the variability of full vaccine coverage at the individual-level (fixed effects) and also explaining the regional-level variation by random intercepts, but the regression slopes are assumed fixed. We adopted the forward selection strategy where factors were added step by step at each model through evaluating their effects by p-values and fit index (log likelihood). In order to best fit the models or in the absence of improved log likelihood values, predictors with a p-value greater than 0.2 were dropped from analysis. Secondly, we developed random-slope modes that may explain variability at the regional-level. For the simplicity of model estimation, each predictor was added one by one in the random component of models. And also, nominal variables (e.g. where the child was delivered and who assisted the delivery of child) were dictomized. Effects of predictors were evaluated by coefficients for slopes and changes in log likelihood as compared to the random-intercept model. In general, we conducted a step-by-step generation of a basic model using a model contrasting approach illustrated by log likelihood values. P-values less than or equal to 0.05 was considered as the level of significance. Maximum likelihood estimates were applied. Statistical analysis was carried out using Stata SE/11 for Windows.

The authors have used data from the Malawi Welfare Monitoring Survey (WMS). The survey was carried out by the National Statistical Office (NSO) in Malawi which followed rules for confidentiality and anonymity. Statistics Norway, through the institutional cooperation with NSO, has given professional input to several stages of the data collection, but neither of the authors have been involved in this work. Statistics Norway has received official permission from NSO to use the dataset for vaccination research. A report on general findings from the WMS is publicly available.

Malawi consists of 27 administrative regions. The variations in vaccination coverage are displayed in Table 1, which also documented regional variations in deliveries assisted by trained personnel, and percentage of the population living in permanent dwellings. If we exclude Likoma, a small group of islands in Lake Malawi and with only 21 observations in the WMS, the highest ranked region was Nsanje, with coverage of 74%. Seven regions had coverage over 50%. The four regions with the lowest ranking had less than 10%vaccine coverage. Although there is no clear pattern to where the regions with high and low vaccine coverage are situated, many of the districts in the Northern Region are among those with low vaccination rates (see Figure 1). A closer inspection of the data revealed that the regional variation displayed in Table 1 were robust to changes in age groups and the exclusion of children without vaccination cards. We will therefore include all children ages 10–60 months in our study, since the number of observations is highest in this case.

Although the full vaccination rate is very low in some regions, the rates are substantially higher for many of the single vaccines (Link to Additional file 1: Table S1 detailing coverage for each separate vaccine).

In Table 2, descriptive statistics with information on vaccination uptake are presented for all the variables included in the analysis. Variations related to gender, family typology and education appear to be small or non-existent. A minor gradient seems to exist regarding household consumption per capita, in which the poorer households have lower coverage. Children born in 2006 naturally have markedly lower coverage since many of the children have not yet filled one year and have therefore not yet received all vaccinations.

Variables describing the circumstances around the delivery of the child seem to have a modest impact on coverage. In general, coverage is higher for mothers that used more specialized health care facilities during birth, and for mothers that had qualified health personnel present during birth. More surprisingly, walking distances to health care facilities do not seem to have an impact on vaccine coverage.

In Table 3, we presented a series of random intercept models which included predictors for the full vaccine coverage at the individual-level and the regional-level. The final fitted models were only presented in Table 3. Details about the analysis are available from authors upon request. In Model 1, child’s age and gender were included as predictors for the full vaccine coverage. Children born in 2006 had significantly lower odds to be fully vaccinated compared to those born in 2002. Model 2 included maternal factors, suggesting that mothers who could not read and write a simple sentence had a significantly lower probability to fully vaccinate their child. As to predictors related to the household income and conditions (Model 3), the model indicated that children in households with a higher annual income (those in the 3^rd and 4^th quintiles) had a significantly higher likelihood to be fully vaccinated as compared to those children in the lowest quintile group. Similarly, households with unsafe water source and without toilet facilities were significant predictors for not being fully vaccinated. Since gender (p= 0.769) and marital status (p=0.844) had no significant effects, we dropped it from further analysis.

In Model 4, factors related to health care were examined by including the place of delivery, who assisted the delivery of child, access (time to walk) to nearest health institutions and the level of trust to hospital staffs. Children born in health centre and health post had significantly lower odds to be fully vaccinated compared to those who were born in maternity clinics. Moreover, there was a higher chance of being fully vaccinated if the delivery was assisted by midwifes or nurses than if the delivery was assisted by doctors or clinical officers. In contrast, the delivery assisted by mothers themselves had a significantly lower probability to be fully vaccinated. Factors related to access (time to walk) to nearest health institutions and the level of trust to hospital staff had no significant effects, but retained in the further analyses due to improved log likelihood values.

In Model 5, the regional-level factors, proportion of delivery attended by health personnel and proportion of permanent dwelling were added as continuous variables and neither of them had a significant effect. In general, predictors in the final model (Model 5) also resulted in considerable improvements of the log likelihood, reflecting the superior fit of this model compared to the earlier models – factors in the fitted model have a significant contribution in explaining the vaccine coverage among the study population. Furthermore, we developed the caterpillar plot (Additional file 2: Figure S1) using random intercept estimates and standard errors from the Model 5. As shown in Additional file 2: Figure S1, the residual plots of 26 regions, one for each region, support the between regions variability for the vaccination coverage, and also indicate a necessity for the random slope model.

In the above mentioned models, we only included random intercepts, so that our next models were primarily developed to explain the variation of full vaccine coverage at the regional-level by a random slope. For this purpose, we continued to fit a random slope in the final fitted model in Table 3 (Model 5), and presented the random component and fit index in Table 4. These include the individual-level factors such as household income, maternal education, where the child was delivered and who assisted the delivery of child. Except for self-attended delivery, those who born in health centre/post and home, and maternal education; the random-slope models for other predictors have considerable improvements in the log likelihood values, reflecting the superior fit of these random-slope models in predicting the variability of full-vaccine coverage across regions compared to the nested random-intercept model. Moreover, most parameter estimates are much larger than the corresponding standard errors, showing the significance (p<0.05) effects of random slopes: in regions with larger coefficients for household income, delivery attended by doctors and midwife, and child delivery at hospitals/maternity clinic slope, these factors have a larger impact on full-vaccine coverage, and vice versa.