Gadgets

Ecological and individual effects in childhood immunisation uptake: A multi-level approach

Description
Ecological and individual effects in childhood immunisation uptake: A multi-level approach
Categories
Published
of 8
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  Sot. Sci. Med. Vol. 33, No. 4, pp. 501-508. 199 I Printed in Great Britain. All rights reserved 0277-9536/91 3.00 + 0.00 Copyright Q 1991 Pergamon Press plc ECOLOGICAL AND INDIVIDUAL EFFECTS IN CHILDHOOD IMMUNISATION UPTAKE: A MULTI-LEVEL APPROACH KJZLVYN JONES’, zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIH RAHAM MOONS nd zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQP NDREW CLEGG’ ‘Department of Geography, Portsmouth Polytechnic, Buckingham Building, Lion Terrace, Portsmouth PO1 3AS, *School of Social and Historical Studies, Portsmouth Polytechnic, Milldam, Burnaby Road, Portsmouth, PO1 3AS and ‘Community and Small Hospitals Unit, Portsmouth and South East Hampshire Health Authority, St Marks House, Derby Road, North End, Portsmouth, U.K. Abstract-Analyses of childhood immunisation uptake have traditionally been conducted at either the ecological or the individual scale. In this paper the problems stemming from these distinct strategies are explored and the potential of a multi-level modelling approach taking simultaneous account of processes at both levels is discussed. This discussion is set in the context of a case-study of pertussis immunisation uptake using data gathered from routine child health surveillance and immunisation uptake monitoring. The role of multi-level modelling in medical geographic research is briefly evaluated. zyxwvutsrqponmlkjihgfedcbaZ Key worrls-immunisation, United Kingdom, geographical variation, logistic regression, multi-level INTRODUmION A wide-ranging analysis of the health issues facing Britain in the 199Os, notes that “(Britain) compares poorly with other countries, especially Scandinavia and the U.S.A., in uptake levels for vaccination against pertussis (whooping cough). . .” 11. In 1989 the European Regional Office of the WHO reported that the majority of its member countries had uptake rates above 80 . Britain only achieved 73 [2,3]. With some 7 of the target population being contra- indicated or otherwise excluded, much higher uptake levels are clearly attainable [4]. Such national figures mask considerable local variations. Studies in Northumberland Health Authority indicate that, though almost 50 of practices have pertussis immunisation uptake rates above 90 , some 20 achieve rates below 80 [5]. In Portsmouth and S. E. Hampshire Health District ward-level uptake rates range between 40 and 85 (61. Such variation is suggestive of a clear medical geography of immunis- ation uptake. The derivation of an explanation for that geography forms the subject matter of this paper. EXISTING EXPLANATIONS FOR PERTUSSIS IMMUNISATION UPTAKE At root, the geographical variation evident in pertussis immunisation uptake in Britain is a conse- quence of its voluntary nature. The decision to immunise or not is essentially a matter of choice; there is no compulsion although there may be con- straints. The geography of uptake is therefore, to an extent, a geography of the ability or inclination to participate in preventive health care and the con- straining factors which limit that ability or incli- nation. Two explanatory themes emerge from the existing literature on uptake: parental choices and professional advice. zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA Parental choices In practice choices regarding an immunisation which should take place by the age of two fall, not to the actual consumer, but to the parent. These choices are not freely made and can be analysed within three frames of reference: time-space constraints, the depri- vation context and fears over safety of the vaccine. Time-space constraints may serve to limit a parent’s access to immunisation services. Such ser- vices may suffer from poor physical accessibility, they may be less easily accessible to people without cars, and they may be open only at inconvenient times. Senior et al. suggest that these issues are likely to be exacerbated by gender role constraints [7]. Women, ascribed the main role in childcare, tend to have reduced access to private transport and less flexible work arrangements. The deprivation context to parental choices regard- ing immunisation links to well-established theses concerning inequalities in preventive care utilisation (81. A classic panoply of social indicators have been linked to poor uptake. These include overcrowding [9], local authority housing tenure [lo] and high population mobility [I 11. The last is a particularly important factor due to its implications for the effectiveness of call-recall systems for notifying im- munisation appointments. Less clear cut associations have been found be- tween uptake and social class (12,131 and ethnicity (14-161. In the latter case it appears, contrary to traditional theories of disadvantage, that ethnic min- orities have higher uptake. This is seen as an indirect outcome of communication barriers preventing dis- semination of knowledge via the media about poss- ible serious side-effects, coupled with respect for health professionals. Knowledge of side-effects also relates to the social class association. Debates focus on whether higher social status, as traditionally ex- pected, produces greater uptake, or whether the zyxwvutsr 501  5 2 ~vYNJomSel zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPON l interlinkage of education and social status leads to an enhanced concern about side effects and hence lower uptake. Fears over the safety of the killed whole ceil Bordetella pertussis vaccine [17] have clearly had an effect on uptake levels. National uptake rates fell from around 80 in 1973 to less than 40 in 1976 [3]. The National Childhood Encephalopathy Study has since indicated that the risks from the disease exceed the risks of immunisation. It is thought that about 40 of parents still harbour worries concern- ing side-effects [18, 191 and that such fears particu- larly affect mothers [20]. It is possible that the persistence of these worries has resulted in persisting non-compliance in some familes [21]. Professional advice The health professional is a key figure in mediating information to the parent regarding the safety of immunisation, its purpose and the attendent sched- ules. Uptake rates will reflect the relative success of this mediation process. The effectiveness of lay-pro- fessional communication provides a backdrop to this issue [22], however specific concern has tended to focus on issues of service organisation and the pro- fessional knowledge base. The impact of service organisation has been con- sidered in two related ways. First, suggestions have been made that commitment and organisation on the part of an individual general practitioner, her/his practice, and the local family health services organis- ation or district health authority may be central in promoting high uptake [23,24]. Generally commit- ment and organisation are held to be synonymous with computerised call-recall systems. Second, cer- tain forms of practice organisation have been found to be associated with low uptake rates. There is thought to be poorer uptake in health authority children’s clinics in comparison with general prac- titioner figures [3] while Jarman et al. claim that older and single-handed general practitioners return low uptake rates [9]. Sound knowledge regarding immunisation is clearly central to the provision of accurate profe- sional advice. Harding and Bolden argue that, despite the rise of “popular medicine”, general practitioners, health visitors and practice nurses are still the main source of information on immunisation [21]. Yet up to 50 of general practitioners are thought to have inaccurate knowledge of the contra-indications and side-effects of pertussis vaccine and lack up-to-date information on its safety [25]. Indeed some general practitioners remain less than wholehearted in their support for pertussis immunisation campaigns [26] either through doubts over safety or because of uncertainties about the effectiveness of the vaccine 1271. Levels of analysis Missing from this review is any discussion of the level of analysis. Studies have been conducted at either the individual or the ecological level. Some refer to factors predicting the probability of an individual being immunised, while others refer to the ‘determinants’ of ward or treatment centre uptake levels in an ecological analysis. Individual level analyses are problematic in that they require access to individual-level data. This access may be refused on grounds of confidentiality. Analytical requirements focus on logistic or log linear modelling and demand considerable expertise. There are consequently very few individual-level explo- rations of immunisation uptake [28]. Ecological analyses, in contrast, are relatively more common (e.g. Refs [9,29]) although they have been recognised as providing “very questionable clues to immunis- ation behaviour” [7]. Both types of study are limited. An individual-level study may commit the atomistic fallacy while aggre- gate studies risk the ecological fallacy [30]. The level of analysis seldom gets the attention that it deserves; at different levels different processes may be operat- ing. Furthermore, there may be an interaction be- tween the levels such that an effect at one level may be constrained or mediated by an effect at another. In the case under discussion in this paper, the individual level processes are those identified above under the heading parental choice. At the ecological level they refer to professional advice provided at treatment centres. The remainder of this paper is directed to an exploration of individual and ecological predictors of uptake in the unified framework of multi-level mod- elling. The data to be used in the study will be described in the next section. Attention will then shift to a consideration of multi-level modelling and its application to the case of pertussis immunisation uptake in one health authority. The paper will con- clude with an assessment of the broader implications of the multi-level approach for quantitative medical geography. CASE STUDY DATA In order to explore the contentions outlined in the previous section data are required on both individual parental characteristics and the nature of the service available at a treatment centre. Such data need to be set against information on pertussis immunisation uptake. Uptake data are routinely collected in all Dis- trict Health Authorities. In those where the collec- tion process is integrated with a call-recall system of immunisation notification, individual-level data should be readily available subject to confidentiality. Data on treatment centre characteristics should be equally easy to obtain from either a District Health Authority or a Family Health Services Authority. Rather more problematic are data on client charac- teristics which may constitute statistical predictors in a model of immunisation uptake. Such data, with the necessary informational detail, can usually only be obtained in those authorities which operate birth/child health surveillance systems. In the study reported here the necessary data are for one District Health Authority. Two data sets were obtained and were linked for subsequent analysis [31]. The first comprised a year’s birth cohort of 6000 children who should have completed their pertussis immunisation programme. It contained the sex, birthday and immunisation status of the child to- gether with the postcode of their current place of  Ecological and individual effects in childhood immunisation uptake 503 Table 1 Variable Previous infant death in family Smoking mother Tenure Stability of parental relations Employment sector Mother’s age Treatment centre type Rationale for selection May lead a parent to be particularly positive towards preventive health care. The variable dichotomises families who have experienced an infant death in the farnly from those who have not. A surrogate indicator for social status 139. This indicator identifies smokers and non-smokers. The expectation is that the children of mothers who smoke will have lower uptake. Related at an ecological level to uptake rates [IO]. The indicator distinguishes owner-occupiers. local authority tenants, tenants in the private rented sector and residents in Ministry of Defence married quarters. It is expected that families who do not own their own home will have lower uptake. An unstable relationship might inhibit uptake through the operation of gender role and time-space constraints limiting parental access to immunisation services. This variable uses the health visitor’s views to classify a partnership as stable, unstable or uncertain. Assesses the impact of membership of key employment sectors. The sectors identified are civilian employment, the military and unemployment. The latter two might be expected to have lower uptakes; the military because of their high mobility, and unemployed people as a consequence of general social deprivation [I I, 121. Several alternative hypotheses: younger mothers may be less likely to have their children immunised; alternatively they may be more likely to bow to professional pressure to immunise. Older mothers may be more aware of negative publicity about pertussis immunisation. Evidence suggests that general practitioner surgeries return higher uptake rates than local authority child health clinics. General practitioners who do not participate in a computerised call-recall system might be expected to have lower rates. These three categories are employed to define this variable residence and a code for the centre where their immunisations should have taken place. This latter code was used to separate general practitioner surgeries, health-authority child clinics and non- participating general practices which did not use the computerised call-recall system. The second data set contained over 100 items of information recorded at birth and on a subsequent visit by a health visitor. An eight month period was in common for the two data sets. The linkage of the two data sets was undertaken using the BMDP Data Manager package [32] with matching taking place on three common data el- ements: birthday, postcode and sex. It is contended that the likelihood of mismatching with 6500 births divided roughly equally between the sexes and among some 19,000 postcodes is low; the chances of getting two of the 6500 individuals with the same sex, birthday and postcode are less than five in ten thousand (311. The resultant linked data set excludes all children registered on one database but not both. Thus all children who moved into, out of or within the Health District between the time of birth and completion of the immunisation schedule are ex- cluded as their postcodes would not match. Coding and transcription errors also affected this process. The final linked data set consisted of 2048 children and combines immunisation status, social infor- mation concerning carer(s), and treatment-centre membership and classification. An exhaustive analy- sis comparing matched and unmatched children by immunisation status did not reveal any substantial bias (311. The choice of variables for further analysis was informed by the research outlined in the previous section and available data. Seven potential explana- tory variables were selected, see Table 1. These data were abstracted from the linked data set and form a two-level data set. At level 1 there are 2048 children; these are ‘nested’ into 126 treat- ment centres at level 2. The dependent variable in the model is the categorical outcome immunised or not. The predictor variables at level 1 are the individual parent characteristics. Nearly all of these are also categorical, the exception being mother’s age. A single level 2 predictor is used: the type of treatment centre which the child attends. Table 2 presents the relevant summary data disaggregated, where appropriate, for the classes of each variable. Evidence from the Census and other surveys suggests that the data are repre- sentative of the local area [31]. The data set provides the basic input for a multi-level model of pertussis uptake. MULTI-LEVEL MODELS The hierarchical, two-level data structure suggests that a multi-level model is required for effective analysis [34]. The object of statistical modelling is to provide a simplified representation of the underlying population. This is achieved by separating systematic features of the data from random variation. Consider the simplest bivariate case involving two continuous variables. The individual-level model can be specified as: Table 2. Variables used in the analysis Individual level Immunisation uptake Previous infant death Smoking mother Tenure: Owner occupied Married quarters local authority Private rented Parental relations: Stable Unsure Not stable Employment status: Civilian Military service Unemployed Mother’s age Level 2: treatment centre Centre type: Participating GP Clinic Non-participating GP 2048 children and 126 centres. Percent Mean 69 3 30 63 a 23 6 64 33 3 70 19 II 28 74 16 10  504 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA ELVYN JONES t al. The response variable (pV) is the probability of child i in treatment centre j being immunised, while the single predictor (x,) might be mother’s age. The systematic component in this model consists of two fixed parameters: the intercept, PO, and the slope, fi, . The random part consists of the residuals, cii. Making the usual assumptions that these are nor- mally distributed with a mean of zero and that there is constant variability (homoscedasticity) and no autocorrelation, the random variation can then be captured by a single measure, ~3, the variance or the ‘average size’ of the random variation. However this model ignores the higher treatment-centre level in postulating that the same relationship between mother’s age and immunisation exists for all centres (Fig. la). In contrast, Fig. l(b) consists of a set of parallel lines, one for each centre, which have different inter- cepts. These random, allowed-to-vary intercepts per- mit different centres to have different levels of uptake around the overall, fixed average for all centres. This is achieved by re-specifying the child-level model of equation (1): to allow the intercept term joj to vary in a centre-level model: BOj = BO Pj The uj terms are the centre-level random terms at level 2; making the same assumptions as for the (0 ) Fired relationships _ (b) Random intercepts I C) Random sloven and intercoats Mothers age (x,, 1 Fig. 1. Alternative models. level-l random part, they can be summarised by a single variance term, crt. In effect, such a formulation allows for observations within centres to be non-inde- pendent, that is autocorrelated [34] with clients in a treatment centre being more alike than a random sample. Finally, Fig. l(c) allows both intercepts and slopes for each centre to vary around the overall fixed average for all centres. Potentially, this allows differ- ing relationships between uptake and mother’s age in different centres; there may be centres where there is little or no relationship, while in others it may be quite steep. Such a fully random two-level model is specified by allowing the slope term of the child-level model: Pg = B0, + Bljx, + Ev (4) to vary according to a further treatment-centre model: Blj= Bi + rj 5) The I’, are another set of level-2 random terms. Making the usual assumptions, these can be summar- ised in a single variance term, gt. It is possible that the slopes and intercepts are correlated so that, for example, centres with a high probability of uptake are associated with a less steep marginal increase in the age/uptake relationship. Conceptually this implies that some centres may have high uptake despite the characteristics of their clients, and vice zyxwvutsrqponmlkjihg ersa This is allowed for by a covariance term between the two higher-level random terms, op,., To summarise, if on estimation ui is found to be small in relation to sampling error, there is no tendency for centres to vary in their uptake once age is taken into account; if et is small, age-uptake relationships do not vary greatly from centre to centre; if a,, is small, variations in the uptake-age relationship are unrelated to centre average uptake. If all three terms are small, multi-level models are not needed, and the single-level equation of (1) is appro- priate. Any apparent centre variation is merely the result of the centre’s composition in terms of mother’s age. The models can be extended [35] by including further fixed terms in the child-level model, by incor- porating centre-level predictors such as centre type in the centre-level model, by allowing the additional child-level predictors to vary at the centre level, and by incorporating predictors at an even higher level (e.g. health authority expenditures). Estimation with binary response Having dealt with the multi-level approach in general terms, attention now focusses on the specific problems of modelling individual immunisation up- take. In this case the response variable is binary not continuous. While this can be modelled by treating the response as the probability of being immunised, this formulation results in three problems [36]: (1) nonsensical values: while the actual re- sponse is either a 1 for immunised or 0 for not, the fitted values can lie outside this range; (2) functional form: a linear relationship be- tween the response and the predictor variables is inappropriate as the ‘effort’ required to raise the  Ecological and individual effects in childhood immunisation uptake 505 proportion immunised from 0.99 to 1 is greater than moving it from 0.50 to 0.51; (3) variance heterogeneity: the assumption of a homoscedastic child-level random part is not appropriate when the actual values of the re- sponse variable are discrete. Usually these problems are overcome by using the framework of generalised linear models (GLMs) [37] which allows the separate specification of a link function between the response and the predictor variables, and a specific distribution for the error term. Discrete responses are accommodated by speci- fying a logit link function and a binomial error term. Recently, the GLM framework has been extended to include multi-level models [38]. A bivariate random-intercept logit model can be specified as a child-level model: W,) = Ptj = exp 8Oj + Bhj~ 1 + eXP j + BAjl (6) and an associated centre-level model: BOj = BO zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA j Equation (7) is equivalent to equation (3); while in equation (6) the yil is the observed response of 1 or 0 for the ith child attending the jth centre and plj is a theoretical entity, the probability of being immu- nised. According to this model, the probability of being immunised varies in a fixed relationship with the predictor (xii) and is allowed to vary from centre to centre according to the random intercept, S,. The child-level model is non-linear thereby resolvmg the problem of appropriate functional form. For esti- mation it can be linearized by taking a logit trans- formation of both sides of the equation: L”g=](P,l(l zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA -P, l= BOj + Blxg (8) The entity p,/(l -pu) represents the odds of an individual being immunised and ranges from zero to infinity. The natural logarithm of this value, the logit, ranges from minus to plus infinity as pii ranges from zero to one. As the logit cannot exceed infinity, pii must be bounded by zero and one, thereby overcom- ing the problem of nonsensical values. The estimated slope parameter of equation (7) therefore represents the change in the log-odds of being immunised as the value of the predictor increases by one unit. As discussed by Wong and Mason [39] computational costs can be reduced without altering interpretation by grouping together observations with the same child-level values. The child-level then becomes not individual children but children in cell i with unique values on the predictor variables who are attached to clinic j. When all the predictors of Table 2 are used there are 1767 such cells implying that the majority of cells consist of a single child. As with single-level GLM’s the variance heterogeneity problem is accom- modated by assuming a binomial distribution for the child-level random part. TO calibrate the model the procedure of Longford [38] has been used. The associated software 1401 is easy to use and can deal with large data sets. The child-level cell random term is not identified and is constrained to unity as the model has a binomial distribution. A limitation of the package is that it has no facilities for data handling and display, conse- quently BMD-P was used for pre-processing and MINITAB for graphing of residuals [32,41]. Imbalance A potential problem with the analysis of this particular data set is that of ‘imbalance’. Given that there are 2048 children and 126 centres there is an imputed average of 16 children per centre. Because of the ‘uncontrolled’ nature of data acquisition there are, in fact, seven centres with more than forty children and five with only one child. Moreover, it is unlikely that the children in each centre represent a random sample of all children in the authority; it is likely, for example, due to the geographical distri- bution of clients and centres, that some centres will have more than a fair share of young mothers. Potentially, such imbalance could lead to severe difficulties of estimation. We would have little confi- dence in an estimate of the effect of mother’s age when two of the three mothers attending a particular treatment centre were aged 26, and the other was 27 years old. Such problems of imbalance are accommodated in a multi-level model [35]. The variation for all centres is ‘pooled’ in the estimation of the effect for each centre. Consequently the estimators ‘borrow strength’ from all the data. If there is little infor- mation for a particular centre, that centre’s estimate will be ‘shrunk’ towards the authority-wide estimate which in itself is based on the most reliable centre estimates. In contrast, a reliably estimated centre coefficient is substantially immune to the influence of the relationships that exist in other centres. A number of studies have shown that ‘shrinkage’ estimates represent a considerable improvement over ordinary least squares estimates [42]. zyxwvutsrqponmlkjihgfedcb INTERPRETATION Model A The simplest possible multi-level model that can be fitted to the immunisation data is a null model which consists of only the constant in the fixed part. The estimates for this model are given in Table 3 as are the ratios of the estimates to their standard errors which can be judged in relation to a standard 2 distribution. As a cut-off, if the ratio exceeds 2.0, the estimates are deemed significantly different from zero; although it must be stressed that ‘marginal’ signifi- cance should be treated with care [43]. Remembering that the response variable is now measured on the logit scale, the estimate for the constant (0.86) when untransformed implies that the probability of immunisation for all children across all centres is 69 . (All logits are given with decimal points, all untransformed logits, the probability of being immunised, are given as percentages.) If all the centres had the same uptake as this, the centre-level variance would be close to zero. In fact, the pseudo Z value of 5.2 implies that there are substantial centre differences in uptake around the grand mean of 0.86 with a variance of 0.14. Estimating the residual for each centre allows an appreciation of the magnitude of these differences. The worst centre has a residual
Search
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks