Data
We used a Swedish consultation level dataset for the calendar year 2018, collected for a project on the effects of the use of digital primary care in the Swedish health system (Wilkens, Thulesius et al. 2020). The dataset combines clinical data on traditional officebased and digital contacts by residents in both urban, suburban and rural areas in seven Swedish regions spread across the country. These regions had a total population of 3.2 million, almost a third of Sweden’s 10.2 million population. Because utilization relative to income is the primary factor of study, and child care has different access points and copayment rules in the regions, only patients from age 18 and older were included in the analysis, a total of 726 087 patients.
The sample included all patients who had been diagnosed with at least one of three different types of infections during the study period: urinary tract infections, upper respiratory tract infections, and skin and soft tissue infections. This diagnosisbased selection was made early in the project, because mild infections were the main reasons for contacting digital providers when this service was made available (Ekman, Thulesius et al. 2019). A national identification number enabled linking of data from different sources on the level of each individual. Data on the socio-demographic variables sex, age, income, education, residency, employment status and country of birth of these individuals were collected from Statistics Sweden. Income was defined as the individual’s gross income for the calendar year 2018, hence was neither related to other household income, nor equivalised for household composition. Data on diagnoses from specialised care (outpatient and inpatient) were collected from the National Board of Health and Welfare and used to construct a Charlson comorbidity index (CCI) value for each patient (Charlson, Charlson et al. 2008). The final sample is described in Table 1 and the Venn diagram in Fig. 1.
Table 1
Descriptive statistics of sample
| Office-based contacts only (Group A) | Digital contacts only (Group B) | Both types of contact (Group C) |
Variable | Number | Percentage or SD* | Number | Percentage or SD* | Number | Percentage or SD* |
Number of patients | 607 586 | (83.7%) | 103 264 | (14.2%) | 15 237 | (2.1%) |
Average number of contacts | 4.4 | (6.3) | 1.5 | (1.3) | 6.3 | (7.0) |
Annual gross income, mean SEK | 212 003 | (266 929) | 294 485 | (920 678) | 225 518 | (200 408) |
Age, years, mean (SD) | 54.0 | (20.0) | 35.9 | (12.6) | 35.3 | (13.8) |
Sex | | | | | | |
Women | 341 116 | (56.1%) | 67 494 | (65.4%) | 11 067 | (72.6%) |
Men | 266 47 | (43.9%) | 35 77 | (34.6%) | 4 17 | (27.4%) |
Charlson co-morbidity index groups | | | | | | |
Full health | 517 578 | (85.2%) | 95 036 | (92.0%) | 14 058 | (92.3%) |
One diagnosis | 42 271 | (7.0%) | 3 838 | (3.7%) | 738 | (4.8%) |
Multimorbidity** | 12 008 | (2.0%) | 468 | (0.5%) | 100 | (0.7%) |
Missing | 35 729 | (5.9%) | 3 922 | (3.8%) | 341 | (2.2%) |
Education level | | | | | | |
Elementary | 138 711 | (22.8%) | 8 934 | (8.7%) | 1 702 | (11.2%) |
High school | 275 738 | (45.4%) | 39 665 | (38.4%) | 7 006 | (46.0%) |
University | 184 232 | (30.3%) | 53 97 | (52.3%) | 6 473 | (42.5%) |
Missing | 8 905 | (1.5%) | 695 | (0.7%) | 56 | (0.4%) |
Country of birth | | | | | | |
Foreign | 111 077 | (18.3%) | 15 132 | (14.7%) | 1 675 | (11.0%) |
Sweden | 496 509 | (81.7%) | 88 132 | (85.3%) | 13 562 | (89.0%) |
Geographic region | | | | | | |
Rural | 141 804 | (23.3%) | 4 976 | (4.8%) | 2 467 | (16.2%) |
Sub-urban | 202 07 | (33.3%) | 28 547 | (27.6%) | 4 068 | (26.7%) |
Urban | 263 712 | (43.4%) | 69 741 | (67.5%) | 8 702 | (57.1%) |
Employment status | | | | | | |
Non-employed | 276 990 | (45.6%) | 14 720 | (14.3%) | 2 955 | (19.4%) |
Employed | 330 596 | (54.4%) | 88 544 | (85.7%) | 12 282 | (80.6%) |
* SD, Standard Deviation. Pearson's Chi2-test for the categorical variables and ANOVA for the continuous variables show statistical differences between the three groups (p-value < 0.001) for all variables.
** Individuals with more than one diagnosis recorded in the national patient register.
Source: Digital primary care study; consultation level data for sampling period.
*Groups A, B and C refer to Table 1.
During 2018, 607 586 individuals in the sample regions made at least one officebased visit and no digital contact (group A in Table 1 and Fig. 1), and had on average 4.4 visits. Another 103 264 people (group B) made only digital visits and had on average 1.5 contacts. A smaller group of 15 237 people utilized both types of services (group C). This group had on average 6.3 contacts, of which 4.3 were officebased visits and 2.0 were digital contacts (not shown). In total, 726 087 people had at least one contact with primary care, which is 23% of the population in the seven regions. People in group B (digital contacts only) were considerably younger and had substantially higher income compared with those in group A. Utilization was generally higher among women. The difference was even larger for digital contacts than officebased primary care, mainly due to the infection diagnoses sample with a large group of women with urinary tract infections. The data also show a lower share of patients with co-morbidity, a foreign birth place, residence in rural areas and low education among the digital primary care users.
Methods
The first specific objective, to measure equality in service utilization across income for the two types of primary care contacts, was met by applying a concentration index and a twodimensional graphical illustration with concentration curves. The concentration index (Eq. 1) measures the accumulated utilization of the service across the population ranked from lowest to highest income, defined as two times the covariance of the number of visits (u) and the relative fractional rank of the ith individual in the income distribution (R), divided by the mean of u (𝜇) (Wagstaff, Paci et al. 1991).
Concentration Index = 2cov(ui, Ri)/𝜇. (1)
This results in a value 0 if individuals regardless of income have the same utilization, which in a graphical illustration is represented by the 45-degree line (Fig. 2). A concentration index below 0 means utilization was higher among lowincome individuals and graphs a curve above the diagonal line, a value above 0 describes the opposite situation with prorich utilisation of the studied service.
Concentration indexes were first measured for both types of consultations across the entire sample, i.e., all patients were included in the same income ranking and their utilization by type of service was separately accumulated (results in Fig. 2). This illustrated differences between the two services across income. Next, the same index was applied for two separate groups of patients to measure the distribution within groups. One estimate for patients who had used officebased services at least once (groups A + C in Table 1) and another who had used digital services at least once (groups B + C, results in Fig. 3). Hence, there was an overlap of patients who had used both types of consultations (group C).
To meet the second specific objective, we estimated factors of contribution to the income inequalities in utilization measured per above. The index was decomposed into how much contribution to the income inequality can be attributed to a set of factors (defined by the variables presented in Table 1), by applying a decomposing (Eq. 2) applicable for any linear regression model with a dependent health outcome or utilization variable (y) and independent variables (x) with estimated regression coefficients (β) (Wagstaff, van Doorslaer et al. 2003). The total concentration of utilization, \(C\), with respect to income, is the sum of each factors’ concentration index, \({{{\Sigma }}_{k}C}_{k}\). Further, the weight of each factors’ specific concentration index is given by the elasticity \(\beta\), i.e., each factor’s beta coefficient in the OLS regression. The last term is a residual component, reflecting the income-related inequality in utilization that is not explained by systematic variation in the available factors. In a perfectly specified model this estimate would be zero.
$$C={{\Sigma }}_{k}{(\beta }_{k}\stackrel{-}{x}/\mu ){C}_{k} + {C}_{\epsilon }/\mu$$
2
This decomposing technique proposed by Wagstaff et al in 2003 has been widely used and is transparent in its relatively simple computation and interpretation. We note that it has been criticised for carrying a set of assumptions and that the literature presents many alternative approaches to decomposition. The concentration index is a bivariate rank dependent index, i.e. it relates individuals’ level of utilization to their relative income rank (as applied in this study). Indeed, the applied decomposition method assumes rank ignorability, i.e., it decomposes the utilization part of the covariance and ignores the association between the covariates and rank (Erreygers and Kessels 2013). This means that for any explanation of changes in covariates, the income rank is assumed to remain the same, which may be troublesome as factors that impact utilization often also impact income (Gerdtham, Lundborg et al. 2016). The method also assumes exogeneity, i.e. that the error term is uncorrelated with the independent variables (Heckley, Gerdtham et al. 2016). Hence, our chosen method is descriptive and cannot claim causality even though most literature where its applied uses language like a factor is contributing to a certain part of inequality. However, the several alternatives proposed are, although more sophisticated computationally, also more difficult in their interpretations, as acknowledged by their authors (Erreygers and Kessels 2013, Heckley, Gerdtham et al. 2016).
For the third specific objective, to estimate horizontal equity, we applied indirect standardised utilization for need before relating it to income in a concentration index, which can be either propoor or prorich the same way as crude inequalities. The approach is referred to as the horizontal inequity index, which takes the value zero when individuals with the same need utilize the same volume of care irrespective of their income (Van Doorslaer, Van Der Burg et al. 2000, van Doorslaer, Koolman et al. 2004).
An ordinary least square regression estimates the parameters α, β, and γ (Eq. 3). Utilization was standardised for age, sex and co-morbidity by the needvariables \(\left({x}_{j}\right)\). Further, to avoid biased estimates of the need variables, the nonneed variables \(\left({z}_{k}\right)\) country of birth, education level, employment status and geographic region were used as controlling factors. For example, not controlling for education level could have led to that the model estimated a higher needs adjustment through the correlation with our health status variable, comorbidity. While education is correlated with utilization, we only want standardization of what we have defined as need. The remaining variation should come through the variable income in the estimation of inequity by the concentration index of needsadjusted utilization (van Doorslaer, 2000).
$${y}_{i}= \alpha +{\sum }_{j}{\beta }_{j} {x}_{ji}+ {\sum }_{k}{\gamma }_{k} {z}_{ki}+ \epsilon$$
3
Then the estimated parameters, the individual values of the standardizing variables, and sample means of the controlling variables are used to predict the utilization of each individual \(\left({\widehat{y}}_{i}\right)\)in the sample. Finally, the standardised utilization \(\left({\widehat{y}}_{i}^{is}\right)\)is equal to the individual’s utilization minus the predicted, plus the population average (Eq. 4). This indirect standardisation is the model’s interpretation of needsbased utilization.
$${\widehat{y}}_{i}^{is}={y}_{i}- {\widehat{y}}_{i}+ \stackrel{-}{y}$$
4
Unlike most studies comparing equity for different types of care, commonly to answer if primary care is more or less equitable than hospital services, our two types of contacts are (arguably) intended to meet the same need. Therefore, for the comparison of utilization across the entire sample, we sought to standardise the utilization of both services by the same scale of need and calculated the needs-standardised utilization by total contacts for every patient. Standardising utilisation of the two services separately would have risked that young, well-educated and healthy individuals would have seemed to need relatively more (digital) services than others simply because this is how the service was consumed. After standardization the concentration indexes were calculated separately for all patients who had at least one officebased visit in one group (groups A and C in Table 1), and all patients who had at least one digital visit in another (groups B and C in Table 1). This way, the needs standardisation of each patient’s utilization was the same regardless of what type of visits the patient had, and the income rank was the same, which made the distributions comparable across the same income rank as for the two first crude concentration indexes by Eq. (1). For the comparisons with the income distribution of crude utilization in the groups using only one or the other type of service, we standardised utilization solely on utilization by individuals in the respective groups.