How Do the Determinants of Demand for GP Visits Respond to Higher Supply? An Analysis of Grouped Counts

Teckle, Paulos; Sutton, Matt Article How Do the Determinants of Demand for GP Visits Respond to Higher Supply? An Analysis of Grouped Counts Swiss Journal of Economics and Statistics Provided in Cooperation with: Swiss Society of Economics and Statistics, Zurich Suggested Citation: Teckle, Paulos; Sutton, Matt (2008) : How Do the Determinants of Demand for GP Visits Respond to Higher Supply? An Analysis of Grouped Counts, Swiss Journal of Economics and Statistics, ISSN 2235-6282, Springer, Heidelberg, Vol. 144, Iss. 3, pp. 495-513, https://doi.org/10.1007/BF03399264 This Version is available at: https://hdl.handle.net/10419/185898 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ © Swiss Journal of Economics and Statistics 2008, Vol. 144 (3) 495–513 a Research Scientist, BCCA-Roche Health Economics Fellow, Centre for Health Economics in Cancer, Cancer Control Research Program, Vancouver, BC Canada. b Professor of Health Economics, Health Economics Research Unit, University of Aberdeen, UK. The authors gratefully acknowledge the assistance and expertise of Teresa Bago d’Uva, from Erasmus University Rotterdam, for her assistance in writing the STATA program for grouped counts. They gratefully acknowledge the UK Data Archive, University of Essex, for providing the original data for analysis. Authors bear sole responsibility for the analysis and interpretation. Correspondence: [email protected] How Do the Determinants of Demand for GP Visits Respond to Higher Supply? An Analysis of Grouped Counts Paulos Tecklea and Matt Suttonb JEL-Classification: C51, C42, I10 Keywords: health care utilization: equity: access to services: survey methods 1. Introduction 1.1 Objective The purpose of this paper is to identify significant determinants of the demand for GP visits and to investigate how such demand determinants respond to higher supply. 1.2 Data The primary data source for the analysis is the 1999–2002 Scottish Household Survey (SHS) data collected by a team from Market and Opinion Research International and Taylor Nelson Sofres (formerly NFO Social Research) on behalf of the Scottish Executive. The SHS is a nationally representative sample of 60,806 Scottish households. The first part of the survey collected information including: household composition, housing characteristics, tenure, vehicles available to the household, occupation and type of industry employing the person, highest income in the household and total household income. The second part collected 496 Teckle / Sutton information on an individual basis (random adult) relating housing experiences (including homelessness), qualifications, use of private and public transport, employment status, income from employment and other sources, health measures and GP visits. Dependent Variable The dependent variable is the number of GP visits during the last twelve months. Respondents were asked: “In the last year, approximately how many times have you seen a GP or family doctor about your own health (either at home or at a surgery/ clinic)?” Respondents were provided with six grouped categories: none (1), one or two (2), three to five (3), six to ten (4), more than ten (5) and don’t know (6). Respondents were also asked how convenient it was to use services (including GP services). Exogenous Variables Age, sex and two health indicators including prevalence of any limiting longstanding illness (yes/no) and self-assessed health (measured with an ordinal scale with three possible responses: not good, fairly good, and good), were used as needdeterminants of GP visits. Self-reported measures of health have been found to be a good predictor of a number of health outcomes, such as subsequent use of medical care or mortality (Idler & Ronald, 1990; Idler & Benyamini, 1997; Burstrom & Fredlund, 2001) and morbidity (Idler & Kasl, 1995; Bobak et al., 1998; Lauridsen et al., 2003). We introduced an interaction effect of having a limited long standing illness and poor self-reported general health to capture potential non-linear effects. Accessibility to GP services: The time price of securing GP services, which depends on the travel time to a GP surgery and the wait time in the office, is a major explanatory variable in the model. GP services are free at entry in the UK National Health Service (NHS), but the people cost of visit is incorporated in travel time, cost and waiting time. For the convenience of general practitioner services respondents were asked, “… bearing in mind where they (the general practitioner) are and your own circumstances, please tell me how convenient or inconvenient you would find it to make use of their services during their normal opening hours, assuming you needed to?” The six possible responses provided were, very convenient, fairly convenient, neither convenient nor inconvenient, fairly inconvenient, very inconvenient or no opinion. How Do the Determinants of Demand for GP Visits Respond to Higher Supply? 497 Household income: The survey also collected three household income components including: income from paid employment and/or self-employment, benefits and pensions. This variable summed the total income from the highest income earner and their spouse, where applicable. Income was equivalised (adjusted for household size and composition) using the McClements equivalence scale approach as suggested by Office for National Statistics (Office for National Statistics, 2003). The McClements scale is based on the assumption that smaller working households with the same income as larger households are significantly better off. The scale takes childless, two adult households as the standard (that is, they are weighted by 1) and scales up the income of households with fewer people and scales down the income of households with more. We used the logarithm of equivalised household income in our regression equation. 2. Econometric Analysis Modelling Patterns of Utilisation of Care A number of different econometric methodologies have been proposed in the literature for the modeling of utilisation of GP services (Cameron & Trivedi, 1986; Cameron & Trivedi, 1993; Cameron & Windmeijer, 1996; Jones, 2000; Sarma & Simpson, 2005). Most previous research has analysed GP visits using a binary variable, in which a value of one is used for those who visited their GPs during the period of interest, and zero otherwise (van der Meer et al., 1996; Abasolo et al., 2001). In modeling the demand for GP visits, the nature of the data on utilisation determines the type of econometric methodology employed. The number of GP visits can take non-negative integer count values indicating that conventional ordinary least squares (OLS) estimation techniques are not appropriate (Greene, 1993; Cameron & Trivedi, 1998). In addition the OLS regression assumes the error term to be normally distributed with a possibility of predicting negative values for the dependent variable. Using a count model overcomes these problems by assuming a skewed, discrete distribution and restricting predicted values to non-negative values. The analysis of the GP services utilisation in this study is based on estimation of a generalisation of the Poisson regression model as discussed in Moffatt and Peters (2000). Here the essence of the approach is provided. We started with the most basic count data model, the Poisson model, where the probability of an event, Y, is determined by a Poisson distribution. Consider a discrete random 498 Teckle / Sutton variable Yi representing the GP visit count for individual i. According to the Poisson model, the probability distribution of Yi is given by: PY y e y i i i iy () ! == −λ λ, y = 0,1,2,…,∞ (1) Where it is conventionally assumed that the Poisson mean depends on a vector of explanatory variables xi according to: λβ i xi=exp( ’ ) (2) Where β is a vector of parameters and the first element of the vector, xi, is a constant, so the first element of β is an intercept. One feature of the GP visit data from the Scottish Household Survey requires modification of the simple Poisson process presented above. The actual GP visit count for the respondents was not observed since the response to the question was grouped. Modification to the log-likelihood function was required to account for this. It is also possible that the grouped nature of the dependent variable may have advantages from a sampling perspective by reducing potential misreporting problems where there are higher counts. In other words, respondents who incorrectly recall the actual number of counts might still get it in the correct “group” at high values of the dependent variable. To account for grouping, the set of non-negative integers is partitioned into J mutually exclusive and exhaustive subsets I1,…,IJ, such that each Ij is the set of consecutive integers {aj,aj + 1,…,bj}, with a1 = 0, aj + 1 = bj + 1 for j = 1,2,…, J − 1, and bJ = ∞. The way in which the utilization of GP services question was asked results in knowledge of the set Ij to which the count belongs, but not the count itself. The probability of individual i being in group j is: PY I PY y e yPX jij i i i ji yI iy yI jj () () !( ; ) , ,...∈= == ≡ = ∈∈ ∑∑ −λ λβ, 1 2 JJ (3) Let yi be the realisation of the random variable Yi. We define an indicator dij to take the value of one if yi ∈ Ij, and zero otherwise. Although the y’s are not fully observable, the d’s are, and the log-likelihood function for a sample of size n may be constructed as follows: How Do the Determinants of Demand for GP Visits Respond to Higher Supply? 499 LogL d P Y ij i j j J i n I( ) log[ ( )]β= ∈ == ∑∑ 11 (4) The final group, group J, consists of an infinite number of integers: aJ,aJ + 1,…, ∞. The probability of the count falling in this final group should therefore be expressed as: 1 1 1 − = − ∑Pji j J x( ;)β in order for its evaluation to be possible. The assumption that the mean and variance of the variable in question are equal is an assumption that is too restrictive for many scenarios and makes the Poisson model restrictive and is rarely employed in applied work. The other problem with the GP visits variable from the survey is that the responses are grouped counts and not simple counts. The alternative negative binomial model (NegBin) overcomes these two problems. A common derivation of the negative binomial model is to respecify λ to account for unobserved heterogeneity, a possible source of overdispersion in the dependent variable as follows: λ βε=exp( ’ )exp( )xi Where ε has a gamma distribution with a mean of one and variance α ε can be thought of as either the combined effect of unobserved variables that have been omitted from the model or as another source of pure randomness. The negative binomial probability distribution takes the form: Pr( / ) exp( ( ))( ( )) !, , , ,...Yy yy i i i i i i iy == −=ε λε λε 012 (5) Unlike the Poisson model, the mean and variance of the dependent variable are now allowed to differ as follows: Vy Ey Ey ii i( ) ( )( ( ))=+1α (6) We built our model staring from the basic model that contained demographic variables (age, gender and their interaction) only. We then introduced one socio-economic variable at a time (from economic status, qualification, household income, and material deprivation) and examined how the model performs. Subsequent 500 Teckle / Sutton models were then adjusted for health indicators (self-assessed health and limiting long-standing illness) of individuals. Variables that do not improve the model are excluded. Multivariate analysis using logistic (greater or equal to 3 visits, otherwise zero), ordered logistic and interval regressions were then performed. We performed the grouped negative binomial regression in the final model. The significance of each exogenous dummies compared to the reference group was tested using a t-test. The first part of the paper deals with determinants of GP visits. The second part deals with determinants of poor access to GP visits, influence of lower supply (poor access) on mean GP visits, and estimates the effect of demand factors on GP visits for those with high and low supply. To undertake this we stratified the sample into two groups based on a proxy supply indicator. These are respondents who tend to report good access (very and fairly convenient access) and those with poor access (very and fairly inconvenient access) to GP services. First we ran a logistic regression of having inconvenient access to services, otherwise zero, on socio-demographic and health variables to identify the characteristics of individuals who are more likely to report inconvenient access to services. We used the stratified sample to examine how the determinants of demand for GP visits respond to low and high supply of GPs compared to the rest of the population. To examine whether access influences the determinants of GP visits, grouped negative binomial regression model was run for each access group. 3. Results and Discussion We begin by discussing results on the determinants of GP visits followed by the effects of lower supply on mean GP visits and comparing models with low and high supply of GP services. 3.1 Determinants of GP Visits The nature of the distribution of GP visits indicated that a large proportion of observations (33%) clustered at 1–2 GP visits; and 13% of respondents reported having made more than ten visits in the previous year (Table 1). In all survey years more than eighty percent of the survey respondents reported convenient and about 14% reported inconvenient access to GP services (Table 2). Table 3 presents the alternative estimation results for number of physician visits. We run seven multivariate regression models. These are logistic regression How Do the Determinants of Demand for GP Visits Respond to Higher Supply? 501 Table 1: Summary Stat for the Whole Sample and by Access Groups All respondents Convenient access sample Variable N Mean N Mean GP visits (0) 56865 0.199 45978 0.198 GP visits (1–2) 56865 0.333 45978 0.340 GP visits (3–5) 56865 0.211 45978 0.214 GP visits (6–10) 56865 0.128 45978 0.126 GP visits (11 + )56865 0.126 45978 0.120 Demographic characteristics Age (sd) 58253 49.93(18.6) 46023 49.86(18.3) female 57869 0.570 46057 0.572 Health measures SAH-good 57052 0.519 46096 0.537 SAH-fgood 57052 0.322 46096 0.320 SAH-ngood 57052 0.159 46096 0.143 LLSI 57051 0.249 46095 0.230 Economic status Self employed 57869 0.047 46057 0.048 Part time employment 57869 0.100 46057 0.106 Looking after home/family 57869 0.082 46057 0.084 Permanently retired from work 57869 0.298 46057 0.296 Unemployed and seeking work 57869 0.039 46057 0.040 At school 57869 0.009 46057 0.009 Higher/further education 57869 0.028 46057 0.027 Training scheme 57869 0.002 46057 0.002 Permanently sick or disabled 57869 0.055 46057 0.051 Unable to work 57869 0.009 46057 0.009 Others 57869 0.005 46057 0.005 Owns car 60860 0.662 46096 0.668 Log Household income (sd) 59412 9.488 45010 9.490 Survey year 1999 60866 0.241 46096 0.244 2000 60866 0.255 46096 0.253 2001 60866 0.256 46096 0.258 2002 60866 0.248 46096 0.245 502 Teckle / Sutton Table 1 (continued) All respondents Inconvenient access sample Variable N Mean Min Max GP visits (0) 8662 0.192 0 1 GP visits (1–2) 8662 0.303 0 1 GP visits (3–5) 8662 0.202 0 1 GP visits (6–10) 8662 0.139 0 1 GP visits (11 + )8662 0.161 0 1 Demographic characteristics Age (sd) 8672 51.75(19.6) 15 110 female 8683 0.600 0 1 Health measures SAH-good 8690 0.426 0 1 SAH-fgood 8690 0.333 0 1 SAH-ngood 8690 0.241 0 1 LLSI 8690 0.351 0 1 Economic status Self employed 8683 0.036 0 1 Part time employment 8683 0.077 0 1 Looking after home/family 8683 0.076 0 1 Permanently retired from work 8683 0.327 0 1 Unemployed and seeking work 8683 0.036 0 1 At school 8683 0.007 0 1 Higher/further education 8683 0.023 0 1 Training scheme 8683 0.002 0 1 Permanently sick or disabled 8683 0.076 0 1 Unable to work 8683 0.008 0 1 Others 8683 0.005 0 1 Owns car 8690 0.581 0 1 Log Household income (sd) 8513 9.480 6.44 14.08 Survey year 1999 8690 0.239 0 1 2000 8690 0.272 0 1 2001 8690 0.244 0 1 2002 8690 0.244 0 1 SAH-fgood = self-assessed health fairly good; SAH-ngood = self-assessed health not good; LLSI = Limiting Long standing Illness How Do the Determinants of Demand for GP Visits Respond to Higher Supply? 509 not significant. People with worse reported health and any limiting longstanding illness have higher GP visits. Although the effect of age, gender and health outcome measures on GP visits seem similar for the two access groups (higher versus lower supply group), there is some difference in the size of the effect of economic status. The effect of lower economic status for individuals who reported poor access is much higher than those who reported convenient access. Being unemployed has a very significant effect on GP visits to those who reported low supply. To individuals who have a high supply of GP services, being unemployed is not a significant demand factor. Those in government work/training scheme and reported worse access, for example have significantly higher visits but the same economic status who reported better access have less visits, though not significant. Those unable to work and report inconvenient access have higher odds compared to those who reported convenient access although both of them have higher visits compared to the reference group. Therefore individuals who are unable to work due to illness and have poor access to care are morel likely to see GP (regardless of the access barriers they might have or report). 4. Conclusion The two basic questions we have tried to answer in this paper are: 1) What are the determinants of GP visits in Scotland? 2) How do the determinants of demand for GP visits respond to higher supply? Finding answer may be crucial to improve the health system in general and particularly for resource allocation, to make it more equitable and adjusted to need. Moreover, answer can help to understand causes underlying the spectacular increase of health care public expenditure over the last two decades. In this study, we define health care consumption in terms of GP services utilization and supply of primary care is proxied by convenience to access to GP services. We included variables that represent differences in demand, such as income and economic status. We adjusted for age and gender that are significant predisposing factors. Change in utilization of GP services over time is captured by a series of year dummies with year 1999 as a base year. To address the first question, we run seven regression models and explored how alternative estimation models perform. To answer the second question, we run six regression models. 510 Teckle / Sutton From a health policy perspective, equity of access to health care is a major issue in remote areas. We compared the demand equations for respondents with worse access to GP services to the rest of the population, taking into account differences in demographic and economic characteristics of individuals. Our pooled model demonstrated the partial significance of access and capacity measures in determining utilization. There are limitations in this study, one being that analysis was not undertaken on the process of care (e.g., length of the consultation, waiting times and GP– patient relationships) as the survey did not ask for such information. There was also a lack of information on measures of quality of care. Incorporating evidence– based quality measures (such as in the Quality and Outcomes Framework, in the new General Medical Services contract) could enrich the analysis on the determinants of the demand for GP visits. Another important data issue was the lack of information available on primary care teams. The data used in this study referred to general practitioners only. Information pertaining to the effect of other primary care team members (such as nurses, health visitors, and allied health professionals) on health outcomes and the utilization of primary medical care was not available. An Information Services Division Report from Scotland reviewed patient contact with the primary health care team. It indicated that GP contacts account for a little over half (59%) of all face-to-face contacts, practice nurses at approximately one-fifth (22%), district nurses (13%) and health visitors (6%) (Scottish Executive, 2005). The other important primary medical care activity missing in this analysis was the prescribing behavior of general practitioners From a health policy perspective, our results offer an important and unique insight in that lower supply of (inconvenient access to) health care services affect utilization. Policy makers in Scotland and elsewhere need to recognize the significant negative effects limited access to GP services and morbidity have on health care utilization, in designing primary health care system. References Abasolo, I., R. Manning, and A. M. Jones (2001), “Equity in Utilization of and Access to Public Sector GPs in Spain”, Applied Economics, vol. 33, pp. 349–364. Al Windi, A., E. Dag, and S. Kurt (2002), “The Influence of Perceived WellBeing and Reported Symptoms on Health Care Utilization: A PopulationBased Study”, Journal of Clinical Epidemiology, vol. 55, no. 1, pp. 60–66. How Do the Determinants of Demand for GP Visits Respond to Higher Supply? 511 Bobak, M., H. Pikhart, C. Hertzman, R. Rose, and M. Marmot, (1998), “Socioeconomic Factors, Perceived Control and Self-Reported Health in Russia. A Cross-Sectional Survey”, Social Science & Medicine, vol. 47, no. 2, pp. 269–279. de Boer, A. G. E. M., W. Wijker, and H. C. J. M. de Haes (1997), “Predictors of Health Care Utilization in the Chronically Ill: A Review of the Literature”, Health Policy, vol. 42, no. 2, pp. 101–115. Burstrom, B., and P. Fredlund (2001), “Self Rated Health: Is It as Good a Predictor of Subsequent Mortality among Adults in Lower as Well as in Higher Social Classes?”, Journal of Epidemiology and Community Health, vol. 55, no. 11, pp. 836–840. Cameron, A., and P. Trivedi (1986), “Econometric Models Based on Count Data: Comparisons and Implications of Some Estimators and Tests”, Journal of Applied Economterics, vol. 1, pp. 29–53. Cameron, A., and P. Trivedi (1993), “Tests of Independence in Parametric Models with Applications and Illustrations”, Journal of Business and Economic Statistics, vol. 11, pp. 29–43. Cameron, A., and P. Trivedi (1998), Regression Analysis of Count Data, Cambridge: Cambridge University Press. Cameron, A., and Windmeijer, F. (1996), “R-Squared Measures for Count Data Regression Models with Applications to Health-Care Utilization”, Journal of Business and Economic Statistics, vol. 14, pp. 209–220. Deb, P. and P. Trivedi (1998), “Demand for Medical Care by the Elderly: A Finite Mixture Approach”, Journal of Applied Econometrics, vol. 12, no. 3, pp. 313 –336. Deb, P. and P. K. Trivedi (2002), “The Structure of Demand for Health Care: Latent Class versus Two-Part Models”, Journal of Health Economics, vol. 21, no. 4, pp. 601–625. Fernandez-Olano, C., J. D. L. Hidalgo, R. Cerda-Diaz, M. RequenaGallego, C. Sanchez-Castano, L. Urbistondo-Cascales, and A. Otero-Puime (2006), “Factors Associated with Health Care Utilization by the Elderly in a Public Health Care System”, Health Policy, vol. 75, no. 2, pp. 131–139. Greene, W. (1993), Econometric Analysis, Macmillan Publishing Company, New York. Idler, E. L,. and Y. Benyamini (1997), “Self-Rated Health and Mortality: A Review of Twenty-Seven Community Studies”, Journal of Health and Social Behavior, vol. 38, no. 1, pp. 21–37. 512 Teckle / Sutton Idler, E. L., and S. V. Kasl (1995), “Self-Ratings of Health: Do They also Predict Change in Functional Ability?”, Journal of Gerontology, vol. 50B, pp. S344–S353. Idler, E. L., and A. J. Ronald (1990), “Self-Rated Health and Mortality in the NHANES-I Epidemiologic Follow-Up Study”, American Journal of Public Health, vol. 80, no. 4, p. 446. Jones, A. M. (2000), “Health Econometrics,” in Handbook of Health Economics, J. P. Newhouse, ed., Elsevier North-Holland: Amsterdam, pp. 265–346. Lauridsen, J., T. Christiansen, and U. Hakkienen (2003), “Measuring Inequality in Self-Reported Healthdiscussion of a Recently Suggested Approach Using Finnish Data”, Health Economics, vol. 13, no. 7, pp. 725–732. van der Meer, J. B. W., J. van den Bos, and J. P. Mackenbach (1996), “Socioeconomic Differences in the Utilization of Health Services in a Dutch Population: The Contribution of Health Status”, Health Policy, vol. 37, no. 1, pp. 1–18. Morris, S., M. Sutton and H. Gravelle (2005), “Inequity and Inequality in the Use of Health Care in England: An Empirical Investigation”, Social Science & Medicine, vol. 60, no. 6, pp. 1251–1266. Office for National Statistics (2003), Income Profile and Rent Affordability: New Tenant Working Households 19997/98–2001/02, Joint Centre for Scottish Housing Research, University of St. Andrews and University of Dundee, St. Andrews, Issue no 4. Sarma, S., and W. Simpson (2005), “A Microeconometric Analysis of Canadian Health Care Utilization”, Health Economics, vol. 15, no. 3, pp. 219–239. Scottish Executive (2005), Practice Team Information ISD Scotland, web site http://www.isdscotland.org/general_practice_info. SUMMARY Although there is a substantial literature on the determinants of demand for primary care, few studies have been able to examine how these determinants respond to higher supply. Some demand studies include supply variables or regional dummy variables to allow for different supply conditions. A few have tested for marginal effects of supply variables attributed at a highly aggregated geographic level. However, relatively little is known about whether there is a supply constraint and how demand responses differ across population groups. We used information from a household survey of 60,806 individuals for whom we had detailed information on supply and access conditions. As in many surveys, How Do the Determinants of Demand for GP Visits Respond to Higher Supply? 513 the annual measure of utilisation is a grouped count and we estimate a grouped negative binomial model (NegBin2) of the determinants of demand for general practitioner (GP) visits by Maximum Likelihood. We exploit a variable on which respondents were asked to report the convenience with which they were able to access GP services. We demonstrate the significance of this variable in determining the number of GP visits. We then examine which demand determinants are correlated with reported convenience. Finally, we compare the demand equations for respondents reporting unconstrained access to GPs with respondents reporting constrained access. We find that being unemployed has a significant positive effect on GP visits for individuals who reported poor access. People who own a car and reported a good access to GPs have significantly higher visits.