Gender-specific practice styles and ambulatory health care expenditures

Abstract

This paper explores the role of physician gender in the expenditures for ambulatory care as a potential source of practice style variation. We exploit a large doctor–patient panel dataset based on insurance-claims data from Switzerland to estimate the effect of physician gender on health care expenditures. We find considerable heterogeneity across specialties. In primary care, female doctors are found to produce similar overall expenditures per visit as their male colleagues, but significantly smaller prescribing costs and significantly higher laboratory costs. In secondary-care specialties, we find that women generate lower overall expenditures, which is mainly driven by consultation costs. These findings provide evidence for the existence of sex-specific practice styles that translate into different overall expenditures as well as different compositions of these expenditures.

Appendix

A Gender effects

The gender effect conditional on covariates $W_{ip}$ is identified and given by

$$\begin{aligned} \Delta (W_{ip})&\equiv E[y_{ip}|W_{ip},D_i=1] - E[y_{ip}|W_{ip},D_i=0] \nonumber \\&=g(\alpha + W_{ip} \beta _1+ \beta _2+ W_{ip}\beta _3)-g(\alpha + W_{ip} \beta _1) \end{aligned}$$

(5)

By integrating over the covariate distributions in the population of female physicians, we obtain the unconditional gender effect:

$$\begin{aligned} \Delta \equiv E[\Delta (W_{ip})|D_{i}=1]=E[g(\alpha + W_{ip} \beta _1+ \beta _2+ W_{ip}\beta _3)-g(\alpha + W_{ip} \beta _1)|D_{i}=1]. \end{aligned}$$

(6)

We deliberately choose this definition of the unconditional gender effect because it is related to an interesting “counterfactual experiment” that answers the following question: how would the average outcome of female physicians change if gender is switched from male to female? In fact, the gender effect is analogous to the average treatment effect on the treated, a point which is stressed in the modern literature on wage decompositions [15]. This lends a meaningful interpretation to the gender effect, which then has a clearly defined analogue in a causal analysis framework.

B Further tables and additional estimation results

See Tables 4, 5, 6, 7, 8, 9, 10, 11.

Table 4 Descriptive statistics

Table 5 Distribution of dependent variables, men

Table 6 Distribution of dependent variables, women

Table 7 Marginal effects by gender group: HCE per visit

Table 8 Marginal effects by gender group: consultation costs

Table 9 Marginal effects by gender group: drug costs

Table 10 Marginal effects by gender group: laboratory costs

Table 11 Gender effect on health care expenditures (Poisson QML)

C Illustration of effect heterogeneity

To illustrate effect heterogeneity, we first estimate the conditional gender effect for each observation in the sample, given by ${\hat{\Delta }}(W_{ip})=g({\hat{\alpha }} + W_{ip}{\hat{\beta }}_1 + {\hat{\beta }}_2+ W_{ip}{\hat{\beta }}_3)-g({\hat{\alpha }} + W_{ip}{\hat{\beta }}_1 )$, see Eq. (5). Second, we estimate the kernel density function of ${\hat{\Delta }}(W_{ip})$ using the Epanechnikov kernel. The nonparametric estimation procedure is explained in detail e.g. in ([6], Chapter 9.3).

For the sake of illustration, we only estimate densities of the gender effects for the group of GPs because this is by far the largest sample.

Fig. 1

Distribution of gender effects for GPs (OLS)

Fig. 2

Distribution of gender effects for GPs (PQML)

Figure 1 presents the results for the OLS estimates and Fig. 2 for the PQML estimates. The unconditional (i.e., average) gender effect, as displayed in Tables 3 and 11, is marked with a straight vertical line in each graph. For better readability, we have omitted observations below the first and above the last percentile of each distribution. We note that the gender effect on consultation costs is quite symmetrically distributed around zero, whereas the effect on drug costs is strongly skewed to the left. This means that a small number of female GPs produce sizably smaller drug costs than male GPs with comparable characteristics would produce. A similarly skewed distribution is found for the gender effect on the number of visits. Comparing figures between OLS and PQML estimates, we find that estimated effect heterogeneity—especially for drug costs—is less pronounced in the latter case. On the whole, we conclude that average effects can be the result of considerable heterogeneity in the practice styles of doctors with different characteristics.