Unvaccinated children in years of increasing coverage: how many and who are they? Evidence from 96 low- and middle-income countries

Introduction

Systematic international efforts to provide immunisation against major childhood diseases to all infants began in the late 1970s and early 1980s(Bland & Clements 1998). After rapid increases in coverage during the 1980s, global immunisation coverage remained stable between 1990 and 2000 at rates close to 80%. Since 2000, higher commitment to immunisation at both national and international levels led to a gradual rise in both the availability of new vaccines and in the proportion of children vaccinated (WHO, 2009).

Global achievements, however, mask substantial inter- and intra-country differences (Delamonica et al. 2005; Jones et al. 2009). In 2009, 23.3 million children under 1 year of age did not receive the third dose of Diphtheria-Tetanus-Pertussis vaccine (DTP3); 70% of those in 10 countries: Chad, China, Democratic Republic of the Congo, Ethiopia, India, Indonesia, Kenya, Nigeria, Pakistan and Uganda (WHO, 2012).

Routine vaccination monitoring and research on vaccination uptake tend to report on antigen and dose-specific vaccination rates (i.e. the proportion of children in the target population that have been vaccinated with a specific vaccine) either in terms of coverage (UNICEF, 2005) or timeliness of vaccination (Clark & Sanderson 2009). DTP3 is commonly used because it is delivered only in routine vaccination activities and it reflects the capacity of the system to engage infants in three consecutive vaccination events. Coverage expresses the proportion of targeted children who have received vaccines but does not indicate, for example, the ability of the system to deliver multiple-dose vaccines (Bos & Batson 2000); this is described by measuring the coverage of two doses of the same vaccine (e.g. DTP 1 and 3) and better described by dropout rates (i.e. the proportion of infants who received a dose of a certain vaccine but not a vaccine scheduled for an ulterior age).

One group of children has been strikingly less studied: those who received no doses of any vaccine (‘unvaccinated’)(Smith et al. 2004). This is because the proportion of unvaccinated children cannot be captured in the routine reporting system and it can only be assessed in household surveys (these are children who have never been in contact with the health system, where routine data are generated). In 2007, the WHO Strategic Advisory Group of Experts on Immunization (WHO/SAGE) requested that the WHO’s Department of Immunization, Vaccines and Biologicals undertake a ‘more detailed analysis of children who have not been reached by immunisation services’(WHO, 2008). The objective of this study was to contribute to the understanding of the factors associated with unvaccinated children as defined above by providing countries with a digested information pack on the matter.

Methods

The Demographic and Health Surveys (DHS) and the United Nations’ Children’s Fund (UNICEF) Multiple Indicator Cluster Survey (MICS) are nationally representative, multiple indicator household surveys. In both, probability-based, multi-stage sampling is used to select enumeration areas and households. Caregivers of children younger than 5 years are interviewed to determine children’s immunisation status (DHS Phase III, 1996; UNICEF – Childinfo, 2008).

A total of 263 DHS and MICS surveys with individual subjects’ responses were accessed. Of the 183 DHS (MEASURE-DHS) surveys, 17 were excluded: three had no relevant data for this study, six had restricted access at the time of the analysis, three were sub-national and five had no variables related to vaccination status. Of the 80 MICS surveys [44 MICS2(UNICEF – Child info, 2008) and 36 MICS3(UNICEF – Child info, 2012) datasets], five were excluded: four MICS2 and one MICS3 did not contain vaccination data. MICS1 surveys were not used because datasets were not available. A total of 241 surveys (166 DHS and 75 MICS) were included in the analyses. A list of included and excluded surveys is shown in Table 1 and countries are shown in Figure 1.

Children 12–59 months of age were included in the analyses. Twelve months of age was the lower limit because children of that age would have had the opportunity to receive all routine infant vaccines. The upper limit of 59 months was chosen to ensure a sufficiently large sample to make analyses meaningful.

Vaccines considered for the outcome variables were bacille Calmette-Guérin (BCG), any vaccine containing DTP, oral polio vaccine (OPV) and any vaccine containing measles antigen (MCV). The outcome variable was vaccination status dichotomised as children not having received any vaccination (‘unvaccinated’) vs. children having received at least one dose of any vaccine. A child was labelled as having missing vaccination status if none of the vaccines were documented as either given or not given and excluded from the analyses; as ‘unvaccinated’ when all documented vaccines were recorded as not given; and as having at least one dose, the remainder. The proportion of unvaccinated children was calculated by dividing the number of unvaccinated children by the total number of children with known vaccination status.

A second variable, ‘at least one dose’, was dichotomised as children having received at least one dose of vaccine but not being fully immunised vs. children having received all vaccines. Missing vaccination status was defined and handled as described above. A child was labelled as having had ‘at least one vaccine’ if it had at least one vaccine documented as given but not being fully vaccinated; and as ‘fully vaccinated’ if all eight vaccine doses (1 BCG, 3 DTP, 3 OPV and 1 MCV) were documented as given. Unvaccinated children were excluded. This variable provides and indication of the number and proportion of those children who having had the opportunity to have at least one contact with the vaccination programme could not be fully vaccinated (i.e. a dropout-like indicator).

In DHS and MICS, vaccination status is ascertained either by the date of vaccination recorded in the child health card, by having a mark on the card (a certain code is recorded in the dataset) or by the caregiver’s recall when the child health card was not available or incomplete. We took into account all vaccinations recorded in cards, regardless of the age at vaccination because the focus of these analyses was the access of children to (vaccination) services rather than correctness of vaccination. Compared to vaccinations recorded in cards, caregivers may forget to report a vaccination that was actually administered and documented (Valadez & Weld 1992; Langsten & Hill 1998) or conversely report that a vaccination was given when it was not actually given and not recorded in the card (George et al. 1990). Recall bias may come into play and cause differences in vaccination rates with those children whose caregivers retained the card (Suarez et al. 1997). In this study, a vaccination was considered as given if it was documented by either card or caregiver recall.

The findings of a systematic literature review were used to obtain an initial list of potential predictors. Research articles reporting on routine childhood immunisation were searched in MEDLINE (from 1966), EMBASE (from 1980), The Cochrane Library (last issue), LILACS (Latin American and Caribbean Centre on Health Science Information; 1982), RHINO literature database and the following websites: WHO (comprising WHOLIS; WHO AFRO Vaccine-Preventable Diseases; WHO/AFRO, -PAHO, -SEAR, -Europe, -EMRO, -WPRO Immunization), UNICEF, The GAVI Alliance, MEASURE DHS, The World Bank and Children’s Vaccine’s programme at PATH; and the sites of immunisation programmes of India, China, USA, Nigeria, Indonesia, Brazil, Bangladesh, Pakistan, Ethiopia and RDC. The inclusion criteria were studies on routine vaccinations in children, reporting quantitative coverage data of at least one vaccine. Of the 7784 studies retrieved, 254 studies were included. Reasons for exclusion were duplicate reports, newsletters or editorials, or not focusing on low- and middle-income countries (LMIC). The initial list of potential predictors included age and sex of the child, physical housing characteristics, ethnicity, religion, socio-economic status, place of residence, wealth, area of residence and access indicators, such as distance to health facilities. These were discussed in meetings with WHO and UNICEF staff to obtain a final list for the analyses.

For these analyses, potential predictor variables were dichotomised (values of the predictors in parentheses; the first term in the parentheses represents the value of the potential predictor for the logistic regression analyses): sex of the child (female vs. male), birth order of the child (first birth vs. subsequent births; first birth vs. the second), level of education of the caregiver (lowest level of education vs. all other education levels combined), marital status of caregiver (alone vs. in couple), tetanus toxoid (TT) vaccination status of the caregiver (<2 TT doses vs. two or more TT doses in any pregnancy), in case of child’s illness, decision-making for seeking care or treatment (caregiver does not decide or depends on other partner vs. caregiver decides, in conjunction with the partner or alone), sex of the head of the household (female vs. male), level of education of the caregiver’s partner (lowest level of education vs. all other education levels combined), ethnic and religious group (least common group vs. rest of the groups), number of household members (above the median vs. below the median), number of offspring in the household (above the median vs. below the median), offspring dead (above the median vs. below the median), area of residence (rural vs. urban), radio and television ownership (none vs. yes or more than one), wealth index (poorest vs. each one of the other four quintiles). Table 2 shows the potential predictors of the child being unvaccinated included in this study.

Vaccination and predictor variables were thoroughly searched in all surveys, which had different names and code for the same variables, using an algorithm described elsewhere (Bosch-Capblanch 2011).

Statistical analyses were conducted using STATA/IC 10.0 for Windows (StataCorp, 2007). Coverage estimates with 95% confidence intervals (CI) were produced using the ‘svy’ STATA command to account for the complex survey designs. Odds ratios (OR) representing the likelihood of being unvaccinated for each potential predictor were obtained by simple and multivariable logistic regression analyses. Logistic regression analyses were conducted in the unique or most recent survey for each country.

Results

Numbers and proportions of unvaccinated children

Two hundred and forty-one DHS and MICS surveys were conducted in 96 countries between 1986 and 2007. The total number of children between 12 and 59 months of age in all surveys with known vaccination status was 1 125 574. The overall number of unvaccinated children across all surveys and years was 111 118 (9.9%), and the median proportion of unvaccinated children was 5.3% (inter-quartile range (IQR) 1.9% to 12.4%). Figure 2 shows the distribution of the number of countries by the proportion of unvaccinated children. In the majority of the surveys (56), fewer than 5% of children were unvaccinated; in the remaining countries, the proportion of unvaccinated children ranged from 5.0% to 28.5%.

The proportions of unvaccinated children by country (unique or most recent survey) with 95% confidence intervals are depicted in Figure 3, with countries sorted by the magnitude of the proportion (note that the scales of the x-axes are different in the three bar charts). The 10 countries with the highest proportion of unvaccinated children were Ethiopia (in 2005, 28.5%), Comoros (in 2000, 28.2%), Zimbabwe (in 2005, 27.2%), Lao Peoples’ Democratic Republic (in 2000, 26.6%), Southern Sudan (in 2000, 26.3%), Nigeria (in 2003, 22.6%), Niger (in 2006, 19.9%), Madagascar (in 2004, 19.9%), Central African Republic (in 2000, 17.9%) and Chad (in 2004, 16.7%).

For those countries with more than one survey, we estimated changes in the proportion of unvaccinated children and of children with at least one dose of vaccine (Table 3) comparing the earliest and most recent surveys in each country. 48 countries experienced significant changes: 10 countries reduced the proportion of unvaccinated children with a median annual change of -0.9% (IQR: −1.4% to −0.4%); and in 38 countries, the proportion of unvaccinated children increased with a median change of 0.4% (IQR: 0.2% to 1.4%). 24 countries reduced the proportion of children with at least one dose, in favour of being fully vaccinated. The median annual change was −1% (IQR −1.8% to −0.5%); 24 others increased that proportion (i.e. less fully vaccinated), with a median change of 1.3% (IQR 0.6% to 3%) and 17 others had no significant changes.

Table 3. Number of countries with significant changes in the proportion of unvaccinated and partially vaccinated children

Unvaccinated	Partially vaccinated*
	Better	Worse	Totals
Better	18 (a)	20 (b)	38
Worse	6 (c)	4 (d)	10
Totals	24	24	48

• ^*Letters in parenthesis are related to Figure 4.

The proportion of ‘unvaccinated’, ‘partially vaccinated’ and ‘fully vaccinated’ children can relate to each other in different ways as exemplified using dummy data in Figure 4, where the inner pie represents the baseline proportions arbitrarily set at 33% each, for illustration, and the outer doughnut represents the proportion some time later. In (b), for example, the proportion of unvaccinated children decreases while the proportion of partially vaccinated increases resulting in a smaller proportion of fully vaccinated children (i.e. the improve in non-vaccination leads to a worsening of fully vaccination). In the 48 surveys experiencing significant changes over time in the proportion of unvaccinated and partially vaccinated children, 18 improved in both indicators, 20 in only the proportion of unvaccinated, six in only the proportion of partially vaccinated (Dominican Republic from 1986 to 2007, Ethiopia from 1992 to 1997, Comoros from 1996 to 2000, Kazakhstan from 1995 to 2006, Liberia from 1986 to 2007 and Mali from 1987 to 2006) and 4 worsened in both (Colombia from 1986 to 2005, Kenya from 1989 to 2003, Uganda from 1988 to 2006 and Zimbabwe from 1988 to 2005) (Table 4).

Table 4. Number of surveys according to the odds ratio values (below 1, not significant around one and above one) by predictor

Predictor (reference value)	Simple regression						Total number surveys
	<1		=1		>1
	N	%	N	%	N	%
Birth order – 1st born (vs. 2nd born)	0	0	39	63	23	37	62
Birth order – 1st born (vs. youngest)	2	3	28	45	32	52	62
Education – Least educated	0	0	20	23	66	77	86
Education partner – Least educated	1	2	9	15	51	84	61
Ethnic – Minority groups	10	21	20	42	18	38	48
Household members – More members	6	8	45	58	27	35	78
Marital status - Alone	5	6	70	79	14	16	89
Radio – No	1	1	21	30	49	69	71
Religion – Minority groups	9	16	29	51	19	33	57
Sex – Female	2	2	85	92	5	5	92
Sex head household – Female	11	19	41	71	6	10	58
Sons and daughters dead – More deaths	2	3	21	33	41	64	64
Sons and daughters in household – More	3	3	53	62	30	35	86
Television – No	0	0	31	39	49	61	80
Tetanus before birth – No	0	0	16	23	53	77	69
Wealth index – Poorest (vs. less poor)	5	6	45	53	35	41	85
Wealth index – Poorest (vs. moderately poor)	6	7	33	39	46	54	85
Wealth index – Poorest (vs. rich)	3	3	29	34	54	63	86
Wealth index – Poorest (vs. richest)	3	4	24	28	58	68	85
Child ill decide – No decides alone	0	0	4	13	26	87	30
Residence – Rural	6	7	37	43	43	50	86

• <1 and >1: indicates odds ratios below and above 1, respectively, with confidence intervals not containing the value 1; =1: indicates odds ratios with confidence intervals containing the value 1. The last column has the total number of surveys with data available for each predictor suitable for logistic regression analyses.

Predictors of unvaccinated children

To ascertain the country-specific population characteristics of unvaccinated children and to identify possible entry points for interventions, we produced two types of summaries: (i) country-specific fact sheets containing the proportions of unvaccinated children for each value of the potential predictor variables and the OR describing the association between the potential predictors and the outcome (unvaccinated), one sheet per survey and (ii) for each predictor, OR for all countries were plotted together to illustrate achievements by country. These results are available from the SAGE/WHO website (WHO). The main findings are summarised below.

The distribution of OR (median and inter-quartile ranges) by predictor across surveys is depicted in Figure 5. The median OR (likelihood of being unvaccinated) was greater among the poorest households (as compared with the richest), children with less educated caregiver and caregiver’ partners, children of caregivers unvaccinated against TT and children of caregivers who decide alone regarding the child’s care when the child was ill. Predictors that showed no significant differences were the sex of the child, the sex of the head of the household and the number of household members.

No predictor was associated with being unvaccinated in all surveys. For example, wealth index was significantly associated with being unvaccinated in 58 surveys, 68% of those for which this variable was reported; caregiver’s education in 66 (77%) surveys, partners’ education in 51 (84%), TT vaccination status in 53 (77%) and caregiver deciding when a child is ill in 26 (87%) of surveys (note that not all surveys had data for all predictors). See Table 5 for the number of surveys according to the OR for each predictor.

Table 5. Median odds ratios and inter-quartile ranges across surveys for each predictor (multivariable logistic regression) and both outcomes

	Unvaccinated			At least one dose
	Median	IQR		Median	IQR
Education caregiver – least educated	1.87	1.33	2.87	1.31	1.05	1.67
Education partner – least educated	1.61	1.16	2.52	1.17	1.00	1.44
Tetanus before birth – No	2.53	1.60	3.85	1.36	1.08	1.72
Child ill decision – decides alone	2.19	1.49	3.13	1.33	1.16	1.61
Wealth – poorest (vs.‘less poor’)	1.30	0.98	1.78	1.20	0.99	1.51
Wealth – poorest (vs.‘moderately poor’)	1.79	1.00	2.73	1.34	1.00	1.77
Wealth – poorest (vs.‘rich’)	1.82	1.00	3.09	1.58	1.09	1.95
Wealth – poorest (vs.‘richest’)	2.30	1.04	5.32	1.73	1.12	2.66

Multivariable logistic regression was performed to account for confounding and effect modification. The independent variables were those having the strongest association with the likelihood of being unvaccinated defined as having the highest median OR in the simple logistic regression: education of the caregiver, education of caregiver’s partner, TT vaccination status of the caregiver, decision-making when child is ill and wealth index. Summary results of the multivariable logistic regression are shown in Table 6.

The TT vaccination status of the caregiver was the predictor with the highest association with being unvaccinated (OR 2.53, IQR 1.60 to 3.85). The OR of the wealth index, using the poorest quintile as reference, increased progressively with the other quintiles from the ‘less poor’ (OR 1.30, IQR 0.98 to 1.78) up to the ‘richest’ (OR 2.30, IQR 1.04 to 5.32).

The absolute magnitude of OR for the outcome ‘at least one dose’ was smaller than their equivalents in the ‘unvaccinated’ analysis. The highest OR was observed when comparing the poorest with the richest wealth quintile (OR 1.73, IQR 1.12 to 2.66).

Discussion

Despite steady increases in vaccination coverage over the past decades (WHO, 2009), a significant number of children remain unreached by immunisation services. In responding to WHO/SAGE (WHO/SAGE), we have attempted to provide information on the characteristics of unvaccinated children in a format useful to country immunisation programme managers. Fact sheets were sent to countries as an aid for decision-making. To retain survey-specific information and to avoid giving the false impression that the described associations are global, we have avoided conducting meta-analyses.

It is striking that the study of children not having received a single dose of any vaccine has been relatively neglected by research. A number of countries have had more than 20% children receiving no vaccinations, two of them with large numbers of children under 5 years of age: Nigeria [25 776 000 children in 2010 (United Nations, 2009)] and Ethiopia [13 819 000 children in 2010 (United Nations, 2009)]. While the proportion of unvaccinated children is relatively small in the great majority of countries, there remain children who have had not a single contact with the health system resulting in a vaccination.

Reporting on a single indicator, while being a feasible and timely way to assess the performance of immunisation programmes, does not unveil serious events, such as non-vaccination, because improvements in the coverage of any subset of vaccines do not necessarily entail an increase in fully immunised children or a decrease in the proportion of unvaccinated; the proportion of unvaccinated children can improve while the proportion of fully vaccinated children can be reduced and vice versa. This has implications for performance-based funding schemes as well as programmatic planning, which are often based on a single indicator (GAVI Alliance, 2011). Common measures of immunisation system performance such as antigen-/dose-specific coverage, dropout, proportion of fully immunised and proportion of un-immunised (WHO, 1998; Vandelaer et al. 2008), while related, are actually independent measures. For example, in Ethiopia, DTP3 coverage increased between 2000 and 2005 from 56% to 69% while the proportion of unvaccinated children also increased from 16.7% to 28.5%.

Logistic regression analyses confirm that these children live in the poorest and least well-educated families. The analyses showed that predictors that were frequently and strongly associated with being unvaccinated were limited caregivers’ education, limited caregivers’ partners’ education, poor TT vaccination status of caregiver, poorest household and caregiver deciding alone about the care for the ill child. The association with TT could suggest that services are largely accessible to a sector of the population who is willing to use them, or that households may uptake health services as a whole without distinction of services or that TT immunisation has a positive effect in the subsequent uptake of childhood immunisations. However, household surveys have limited data on health services issues, such as range of activities, staff or other resources, to reach a conclusion.

Both simple and multivariable methods were used to determine the significance and magnitude of the association between potential predictors and the outcome variables. While multivariable analysis is more explanatory and provides a more precise estimates of the contribution of each individual factor associated with being unvaccinated by controlling for the contributions of other factors included in the model, simple logistic regression may be more useful in directing interventions by targeting population characteristics strongly associated with non-vaccination. The ‘diagnostic odds ratio’ has been suggested as a prevalence-independent diagnostic performance indicator (Glas et al. 2003), which allows for comparing tests (in our case, for identifying predictors) and for analysing using logistic regression models. Association with predictors was slightly different when considering unvaccinated children or children with at least one but not all doses of vaccine. Similar findings have been reported elsewhere, although the calculations of partially vaccination rates were not identical to those used here (Smith et al. 2004). Predictors were strongly associated with the fact of being unvaccinated suggesting that these children belong to more extreme situations.

Addressing some of the identified predictors require substantial resources and time; and the impact on vaccination outcomes may not be immediate (e.g. household wealth). However, we purposely included other predictors that could be useful in identifying potential interventions, such as ownership of radio or television (TV) in the household. The absence of radio or TV was strongly associated with an increase in the likelihood of being unvaccinated (in the simple and multivariable logistic regression models) and informs the use of mass media interventions to increase coverage (Grilli et al. 2002).

This analysis had several limitations. First, for some children, the vaccination status was ascertained by caregiver’s recall. A bias may be introduced overall if recall significantly differs between the different predictor groups. Furthermore, the inclusion of children who received vaccines beyond the correct vaccine schedule will have probably reduced the proportion of unvaccinated children. Therefore, our findings should be seen as a best case scenario. Secondly, data for all potential predictors were not available in all surveys. For example, the predictor ‘caregiver’s decision when child is ill’ appeared in only 30 surveys (MEASURE-DHS). Thirdly, DHS and MICS, in their different waves, were designed in slightly different ways. Although data were harmonised prior to the analyses, some inconsistencies may remain undetected. Forth, not all surveys were recent and findings may no longer be relevant in some rapidly changing countries. Finally, many potential predictors of a child receiving no vaccination are likely to be missed by multiple indicator surveys. More targeted surveys enhanced with qualitative methods are likely to provide a more complete picture of the characteristics and causes of a child being unvaccinated.

Conclusion

While routine vaccination coverage monitoring based on specific vaccines provides a feasible and timely way to ascertain the performance of immunisation programmes, serious events (such as being ‘unvaccinated’) and inequities may remain unveiled. Countries' immunisation programmes and national and international immunisation stakeholders should monitor the proportion of unvaccinated children in addition to coverage for specific vaccines. This should be performed periodically or where poor performance is suspected. Nationally representative household surveys provide evidence on those issues and can also be used to ascertain the specific factors that influence access to immunisation services. In our analyses, several factors emerged as important and the country-specific fact sheets made the findings accessible at country level to consider corrective actions.

Acknowledgements

We thank Bernard Brabin, Christian Schindler and Kaspar Wyss for comments on the manuscript, and Jos Vandelaer (UNICEF) for his contributions during the conceptualisation phase. Lise Beck produced Figure 1. K. Banerjee and A. Burton are staff members of the World Health Organization. The authors alone are responsible for the views expressed in this publication and they do not necessarily represent the decisions, policy or views of the World Health Organization.