Socioeconomic determinants of rates of consultation in general practice based on fourth national morbidity survey of general practicesBMJ 1996; 312 doi: https://doi.org/10.1136/bmj.312.7037.1008 (Published 20 April 1996) Cite this as: BMJ 1996;312:1008
- Roy A Carr-Hill, senior research fellowa,
- Nigel Rice, research fellowa,
- Martin Roland, director of research and developmentb
- a Centre for Health Economics, University of York, York YO1 5DD
- b National Primary Care Research and Development Centre, University of Manchester, Manchester M13 9PL
- Correspondence to: Dr Rice.
- Accepted 29 February 1996
Objective: To identify the socioeconomic determinants of consultation rates in general practice.
Design: Analysis of data from the fourth national morbidity survey of general practices (MSGP4) including sociodemographic details of individual patients and small area statistics from the 1991 census. Multilevel modelling techniques were used to take account of both individual patient data and small area statistics to relate socioeconomic and health status factors directly to a measure of general practitioner workload.
Results: Higher rates of consultations were found in patients who were classified as permanently sick, unemployed (especially those who became unemployed during the study year), living in rented accommodation, from the Indian subcontinent, living with a spouse or partner (women only), children living with two parents (girls only), and living in urban areas, especially those living relatively near the practice. When characteristics of individual patients are known and controlled for the role of “indices of deprivation” is considerably reduced. The effect of individual sociodemographic characteristics were shown to vary between different areas.
Conclusions: Demographic and socioeconomic factors can act as powerful predictors of consultation patterns. Though it will always be necessary to retain some local planning discretion, the sets of coefficients estimated for individual level factors, area level characteristics, and for practice groupings may be sufficient to provide an indicative level of demand for general medical services. Although the problems in using socioeconomic data from individual patients would be substantial, these results are relevant to the development of a resource allocation formula for general practice.
Characteristics of individual patients are much more powerful predictors of consulting patterns than the characteristics of the areas in which patients live
The effects of individual socioeconomic factors themselves vary in different geographical areas
Resource allocation methods based on area of residence (for example, Jarman score) will always be inferior to an approach that takes into account the characteristics of individual patients
For both primary and secondary care there is a well established association between deprivation measured in terms of census variables at the area level and various indices of rates of use of general medical services1 2 and hospital and community health services.3 4 Interpretation of this work has been bedevilled by two problems: the extent to which supply of medical services might affect demand5 and the ecological fallacy (the extent to which associations between socioeconomic characteristics of a small area and utilisation rates may not hold for individual patients).6 7
The fourth national morbidity survey of general practice (MSGP4),8 the fourth decennial survey organised by the Royal College of General Practitioners in conjunction with the Office of Population Censuses and Surveys, affords the opportunity to combine an analysis at the individual patient level with data from small areas thereby improving our understanding of the relations between the different needs components, the consultation behaviour of an individual patient, and how these may vary across areas. This complements the analysis of these data already published, which has concentrated on identifying patterns of morbidity rather than workload.8
DETERMINANTS OF USE
The decision to consult a general practitioner is complex; it depends not only on a person's “objective need” to consult (morbidity status) and their propensity to consult given their “objective need” but also on the availability and type9 of services and possibly the characteristics of the local community.10 11 The provision and supply of extra services and staff, however, may potentially induce demand. Estimation of the demand in this “chicken and egg” situation requires appropriate statistical techniques. This was an important consideration in the analysis of the health needs at the small area level for hospital and community health services12 and may be important in a similar analysis at area level for primary care.13 14 This is not a problem here because the way in which an individual patient consults is unlikely to affect overall general practitioner or practice behaviour.
DATA AVAILABLE FOR ANALYSIS
Data were collected by 60 volunteer practices between September 1991 and August 1992 on consultations by all patients fully registered for part or whole of the study period (n=502493). Results are presented here on workload measured by all consultations.15
A socioeconomic questionnaire was administered during the study to all patients on the 60 practice lists regardless of whether or not they had consulted during the study period and was completed by about 83%.8 These questions are a subset of those in the census and provide basic sociodemographic information on age, sex, ethnicity, marital-cohabitation status, whether or not a sole parent, economic position last year and last week, occupation sufficient to generate the registrar general's social class classification, and housing tenure. Because loss of employment16 is seen as important and given the difficulty of empirically disentangling the effects of economic position last week and last year, change in employment status during the study year was also used in the analysis. Whether or not the patient was registered as “permanently sick” was the only direct measure of morbidity. The box shows the codes used in multivariate analysis.
The individual data were supplied together with the following small area statistics from the 1991 census: housing tenure, social class, unemployment status, permanent sickness, student status, car ownership, single carer households, dependent children, elderly living alone, overcrowded households, educational qualifications, and limiting longstanding illness. A rural-urban ward classifier was used, and access variables were also calculated to represent the distance from the centroid of the enumeration district where each patient lived to the centroid of the enumeration district of her or his practice.
For reasons of confidentiality small area statistics were not supplied to us for enumeration districts and wards that were sparsely represented in the study population. In total 1640 districts and 480 wards were identified. Omission of unidentifiable districts resulted in the loss of a third of the sample and introduced an unacceptable level of bias. This was not true for excluded wards and hence these were used as the level for attributing small area census data to individual patients.
For the same reasons, although practice level characteristics were supplied, individual practices were not identified; instead the 60 practices were combined into 19 groups of three or four practices according to the number of doctors per 10000 patients, the number of nurses in the practice, and whether the practice was a training practice or a dispensing practice. Practice effects were modelled as 18 binary dummy variables contrasted against the first grouping, which consisted of three practices with no trainees, no nurses, no dispensing facilities, and relatively small numbers of doctors per 10000 patients.
Only those participating in the study for the full duration of the year were included in the analysis. For better specification six models were estimated separately for males and females aged 0 to 14, 15 to 64, and 65 and over.
The structure of the population modelled was hierarchical with individual patients located within geographical areas (electoral wards of residence) within general practices, but given that only 19 practice groupings were identified, the latter was not used as a separate level in the analysis. Where the data are clustered the assumptions of ordinary least squares regression analysis are not met as the residuals are correlated: parameter estimates of explanatory variables will be biased and SEs underestimated so that confidence intervals will be too narrow. Associations may be reported as significant when they ought not to be.
Multilevel models—which extend linear regression to deal with variability at several levels in the study design—have been developed explicitly to account for clustering of data.17 18 In addition to allowing for variations in the conventional ordinary least squares slope and intercept according to each ward of residence, multilevel modelling also allows for systematic exploration of the ways in which social or geographical characteristics, or both, and practice “style” of an area affect consultation behaviour. This is an important innovation.
Workload is measured by the number of consultations by each patient during the study year. Given the count distribution (for males: 35% had no consultations, 17% one consultation, 5% five consultations, 6% 10 or more consultations15) a Poisson or negative binomial functional form of a multilevel model is appropriate,19 and the latter was chosen after experiments with fitting samples of data. The general form of the multilevel model used is explained elsewhere.17
In a negative binomial regression there is an exponential relation between the dependent variable (the number of consultations an individual patient makes) and the explanatory variables which, therefore, cannot be viewed as simple coefficients of impact. The size of the estimated coefficients together with their associated sign, however, indicate both the relative magnitude of the effect of each variable and the direction of association with the response. Estimation is constrained so that the number of predicted consultations cannot be less than zero.
All individual and census level variables were initially included in the model and a stepwise elimination procedure (P>0.05, t test) was then used. Once a set of fixed individual explanatory variables had been identified, the residuals were analysed to identify any systematic patterns in consultation rates across the wards.
Table 1 gives the breakdown of rates of consultations by the socioeconomic variables in the different age-sex groups. The results are, in general, consistent with previous analyses.8 Table 2 gives the multilevel regression results. The presentation here concentrates on the themes common to all six models.
Patients classified as permanently sick consulted much more frequently. The associated coefficient was much larger than the others in the model particularly among men. The effect was unsurprising given that this was the only direct measure of health status in the dataset.
Change in employment status affected consultation behaviour.20 21 This appeared for both men and women and also for boys (although with smaller magnitude). For adults, the continuously employed (baseline category) consulted the least, while those who had lost employment during the year were generally among the most frequent consulters.16 22 23 Higher rates of consultations were also found among the other employment categories, which may be due to the inclusion of students and the permanently sick within this group.
Ethnicity is associated with differences in consulting patterns.24 25 Among adults, patients of Indian, Pakistani, and Bangladeshi origin consulted most frequently, although, after other variables were controlled for, the effect was significant only for women; a similar pattern was seen for children but in this case, the significant increase was found only among boys. “Other” ethnic minority groups (mostly Chinese) consulted less frequently than their white counterparts for women but more frequently for boys.
Men who smoked consulted more, although the positive association tended to decrease in size with increasing age. Indeed, for elderly people of both sexes smoking was associated with a general decrease in consulting.
Women who were cohabiting (including married) consulted more frequently, mostly because of increased consultations associated with family planning and pregnancy. Among elderly women those cohabiting consulted less frequently, although this effect was restricted to non-owner occupiers.
Distance to practice and rurality both reflect accessibility, and for girls and elderly men particularly, those living nearer consulted more. In addition, women and boys living in urban areas consulted more than their rural counterparts, but again, within the urban area, those living nearer consult more.
Previous studies have shown that children of single mothers are more likely to consult27 28 29 and that children whose parents have divorced experience poorer health than children whose parent has been widowed.30 In these analyses the effect of having a sole parent was found to be significant for girls only, the direction being towards less frequent attendance at the surgery.
Significantly less frequent consultations were also found for fathers with dependent children compared with fathers with no dependent children, though there was no apparent effect for mothers of dependent children. The results do not support previous suggestions on the relation between single parenthood and morbidity.30
Because individual level effects were controlled for, one would not expect many additional area effects to be significant, and indeed only the percentage of households without access to a car and the percentage of owner occupiers appeared significantly in more than one of the six age-sex groups (more detail on these and practice effects are given elsewhere15).
These models provide substantial discrimination, but because of the complex model specification used a r2 type measure is difficult to interpret. An illustration for low, typical, and high consulters among 40 year old women, however, is provided in tables 3 and 4. The predicted rates of consultation vary substantially from 1.3 to 29.0 (table 3). Moreover, changes in single variables have a substantial impact. For example, “high” consulters with all the characteristics defined in column 3 of table 3, except that the head of household is non-manual rather than manual, would have 2.5 fewer expected consultations than a manual equivalent (table 4).
Though there was highly significant variation between areas in each of the six age-sex models, for a typical patient the overall effect is not large, accounting for between 2% and 8% of total unexplained variation. For children and adults, however, the coefficients of age, housing tenure status, and distance from practice were found to vary significantly across electoral wards, and for elderly men the effects of housing tenure varied between areas.15 The results imply, for example, that owner occupiers in one area may have notably more or fewer consultations than owner occupiers with the same socioeconomic and health characteristics belonging to another area. This type of variation may therefore play an important part in determining general practitioner workload, and hence it is important to retain some local planning discretion.
These analyses have identified socioeconomic characteristics measured at both the individual and area level associated with the number of consultations presenting at general practitioner practices sampled across England and Wales. Generally, the results of the models support previous research linking individual “deprivation” characteristics and morbidity, although there are some differences especially in respect of ethnicity and of the “social support” variables such as cohabitation status and lone parenthood. The former may be due to the bias introduced by interviews in the English language and the latter may reflect the decreasing salience of marital status to social position and vulnerability.
While the kind of area variables which appear as significant (for example, rates of car ownership and tenure status) are also similar to those identified in other studies, when modelled correctly in a multilevel framework, there are far fewer significant area effects: reliance on “indices of deprivation” will inevitably be very crude because individual characteristics are so much more powerful predictors of consulting behaviour.
The models described form a basis for estimating the demand for general practitioner services at the practice level. While the problems in using individual socioeconomic data would be substantial, it might be possible, by a combination of small area statistics, the sample of anonymised records, and age-sex registers, to provide a basis for an equitable distribution of resources in the primary as well as the secondary care sector.
We thank Professor Harvey Goldstein, Jon Rasbash, and Min Yang of the Institute of Education for statistical technical advice; and Trevor Sheldon and Peter Smith, University of York for comments on earlier versions of this work.
Funding Department of Health.
Conflict of interest None.