Empirical model
To determine the relation between quality and competition, we estimate a panel data model. We consider the following general linear model:
$$\begin{aligned} y_{it}= \varvec{x}_{it} \varvec{\beta }+v_{it}, \end{aligned}$$
(1)
where
t denotes the year
\((t=1,2,\ldots ,T)\) and
i denotes the hospital (
\(i=1,2,\ldots ,N\)). The independent variables for hospital
i in year
t are given by vector
\(\varvec{x}_{it}\) and the dependent variable is given by
\(y_{it}\). In this model
\(v_{it}=u_{it}+c_{i}\) is the composite error, where
\(c_{i}\) is the unobserved component and
\(u_{it}\) is the idiosyncratic error (see for example [
34]). For each model, we made assumptions on the correlation between the unobserved component
\(c_{i}\) and
\(\varvec{x}_{it}\). If we allow these to be correlated, then we have a fixed-effects model where
\(c_{i}\) is a parameter that we will estimate. If we assume that they are uncorrelated then we have a random-effects model in which we assume a structure for the variance of
\(v_{it}\) (see [
34]).
In our specific application, we estimated a model for each diagnosis. For hospital i in year t \((t=2008,\ldots ,2011)\) we denoted its concentration index by ms (see below for the definition) and its quality by quality. To control for possible case mix differences between hospitals, we included the fraction of females (\(\mathrm{frac}\_\mathrm{female}_{it}\)) and the fraction of patients who are 65 years old or older (\(\mathrm{frac}\_65_{it}\)). To control for the fact that some hospitals deal with one dominant health insurer, while other hospitals deal with several competing insurers, we calculated the HHI of the insurers that a hospital faces (\(\mathrm{HHI}\_\mathrm{ins}_{it}\)). We calculated this for, say, hospital i in year t, by summing up the squared shares that each insurer has in the total revenue of hospital i in year t. Furthermore, given that teaching hospitals may treat more severe patients, we included a dummy variable for teaching hospitals (\(\mathrm{acad}_{it}\)) and since quality may depend on volume, we included a dummy for hospitals with a low volume of patients (\(\mathrm{lowvolume}_{it}\)). In each year, the \(25\%\) of hospitals with the lowest volume were considered as low-volume hospitals.
The estimated model is:
$$\begin{aligned} {\text{quality}}_{{it}} & = \beta _{1} ms_{{it}} + \beta _{2} {\text{frac}}\_{\text{female}}_{{it}} + \beta _{3} {\text{frac}}\_65_{{it}} \\ & + \beta _{4} {\text{HHI}}\_{\text{ins}}_{{it}} + \beta _{5} {\text{lowvolume}}_{{it}} + \beta _{6} {\text{acad}}_{{it}} + v_{{it}} \\ \end{aligned}$$
(2)
To estimate the relationship between quality and competition, we needed to measure market power. There has been a great deal of debate about market definition and the measurement of market power in the literature. For an overview, see [
12]. Many authors use a rather crude measure of competition: for example [
24], measure competition as number of hospitals within a 30-min journey controlled for population density. However, although travel time is an important factor in hospital choice, choices can also be influenced by other patient and hospital characteristics.
Gaynor and Vogt [
12] propose the use of the Logit Competition index (LOCI) to measure market power in the hospital market. The index is based on a weighted average of a hospitals market share per micro-market. The construction of the competition index starts by modeling the demand with a choice model. The choice model includes a utility function which, given characteristics of the consumer and hospital, depends on the utility that a patient derives from each hospital. The utility depends on both observable and non-observable consumer and hospital characteristics. With the logit choice model, it is possible to calculate the probability that a specific consumer type will choose a specific hospital. Each group of patients with similar characteristics (e.g., zip-code, age, gender, diagnosis, etc.) forms a micro market.
Under a standard price competition model, the competition index (LOCI) of hospital
j for consumer type
t is given by (see [
12])
$$\begin{aligned} \Lambda _{j}= \displaystyle \sum _t w_{tj}(1-s_{tj}) \end{aligned}$$
where the weights
\(w_{tj}\) are the relative importance of each consumer type
$$\begin{aligned} w_{tj}= \frac{N_{t} s_{tj}}{\displaystyle \sum _t N_{t} s_{tj}} \end{aligned}$$
and
\(N_t\) is the number of consumers of the type
t.
The LOCI \(\Lambda _{j}\) is a measure of the competitiveness in the market. The index takes on values between 0 and 1, where \(\Lambda = 0\) means that hospital j is monopolist and \(\Lambda = 1\) means that the market is perfectly competitive.
We interpret the LOCI as 1 minus the weighted market share. To simplify the interpretation of our results with we use the variable ms as a shorthand for “market share”, which is constructed as \((1-\Lambda )\).
For our purposes, we are able to use actual market share with the advantage that all non-observable characteristics are implicitly taken into account. Alternatively, we could have used estimated market share, with the advantage that all consumers are taken into account. In our estimations, we used actual market share. However, the use of actual market share could potentially lead to endogeneity: hospitals providing good quality may have higher market share.
Our approach is similar to other articles about the impact of hospital competition on the quality of healthcare, e.g., Gowrisankaran and Town [
13,
14], Kessler and McClellan [
16], Gaynor et al. [
11], and Cooper et al. [
6] all use a market share based on travel distance in their estimations directly or in their IV estimations, in order to avoid potential endogeneity problems. Our approach is different from Forder and Allan [
10]: they use an administrative region (Medium-level Super Output Area) as a market to calculate market share. Since they do not use patient choice models, they cannot rely on market share based on travel distance as IV. They therefore use the level of competition in neighboring areas as IV.
To prevent any endogeneity problems, we also estimate our regressions with an estimated market share (based on travel distance only) in an instrumental variable (IV) approach. Our main contribution is that we define our micro markets at the level of the quality indicator. Our micro markets consisted of the group of DTCs that are linked to the quality indicators. For each quality indicator, we estimated the relationship between the indicator and the competition indicator, which meant that we were able to construct a competition indicator for each quality indicator. The micro markets are defined by a four-digit zip code and diagnosis. The narrower a micro market becomes, the more precise the total market share becomes. However, we should not make our micro markets smaller than four-digit zip codes and diagnosis, because then we would have too few observations per zip code. For example, age is highly skewed for each diagnosis: the cataract and bladder tumor diagnosis groups include mainly elderly patients, while the tonsil diagnosis group consists mainly of younger people. This indicates that splitting the micro markets across age categories will not add a great deal of information. Furthermore, there is no reason to assume that choices would depend on gender.
Quality indicators
For the purposes of this research, we used the quality indicators from the ‘Dutch Healthcare Transparency Program’ (in Dutch: Zichtbare Zorg), which were developed by the Health Inspectorate in order to support various goals such as the provision of information for patients and consumers to help them make their choices, purchase information for health insurers, control information for the Inspectorate and improvement information for providers. The Dutch Healthcare Transparency Program started in 2007 and in 2008 quality indicators became available for ten diagnosis groups. The hospitals are required to provide the registered quality indicators to the government annually.
5 Independent Treatment Centers are not obliged to provide quality indicators, and, for this reason, we have no data on the quality indicators of the Independent Treatment Centers. The quality indicators can be divided into process indicators, structure indicators, and outcome indicators.
Structure indicators relate to the organization and are recorded at the hospital level; process indicators measure the process of activities at the patient level and outcome indicators measure outcome values at the patient level. Although outcome indicators are the most important indicators in terms of informational value, the share of outcome indicators for the Dutch Healthcare Transparency Program is less than 16%. According to an evaluation carried out by the Court of Audit (in Dutch: Algemene Rekenkamer), the indicators of the Dutch Healthcare Transparency Program indicators for the hospital sector are stable and it is therefore possible to analyze trends over the years for which records have been kept [
25].
We used quality indicator data from 2008–2011. Because our time period is relatively short, we used process and structural indicators (with a ratio scale) rather than outcome indicators because process and structure indicators can be influenced by hospital management and not only by medical specialists, and these types of indicators are also used by insurers [
4]. Six selection criteria were applied to include quality indicators for the diagnosis groups in our research sample. (1) The diagnoses are not selected on medical similarity but on whether hospitals can compete for patients. We used quality indicators for the diagnosis groups that made up the market segment. In this segment, the prices for the diagnosis groups are determined by negotiations between insurers and hospitals. This means that hospitals are able to compete on price and quality. This is not the case for all hospital treatments (such as urgent care). The Dutch Healthcare Authority has selected diagnoses for the market segment on criteria such as the transparency of the product definition, price and quality, the existence of market dynamics including entry and exit, the absence of undesirable effects and the absence of high transaction costs [
17]. (2) Hospitals have been obliged to record quality indicators for the diagnosis groups since 2008; however, the number of diagnosis groups has grown over the years. We selected only the quality indicators that have been recorded since 2008. This means that the quality indicators that have been developed by the Health Inspectorate since 2008 are not part of our selection. (3) The quality indicators needed to be comparable over the years. (4) We selected diagnosis groups that involved surgical intervention. (5) We selected high-volume diagnosis groups with over 10,000 treatments per year. (6) We excluded indicators with categorical values (yes or no answer). Our final sample consisted of quality indicators for three diagnosis groups: cataract (ophthalmology), adenoid and tonsils (otolaryngology), and bladder tumor (urology). For bladder tumor, quality indicator data were available for 2008 to 2010 and for cataract, and data were available for 2008 to 2011 for adenoid and tonsils. The three diagnoses are elective care treatments that include daycare surgery. It should be noted that this is also the case for the bladder tumor diagnosis because the quality indicator that was used in this study relates to the low-risk patient group (non-muscle invasive bladder tumor). Table
1 shows the quality indicators that we included in our analysis.
6
Table 1Quality indicators
Cataract |
i02-02 | Complications: percentage of cataract operations without a posterior capsular rupture (including vitrectomy) | Process | 99.63 | 0.35 |
i02-03a | Accurate diagnosis of the second eye: percentage of patients undergoing a cataract operation on both eyes with a gap of more than 28 days between the first and second operation. For a careful assessment of the second eye there should be enough time between the surgery of the first and second eye | Process | 95.56 | 8.40 |
i02-03b | Accurate diagnosis of the second eye: percentage of patients undergoing a cataract operation on both eyes that have (i) a postoperative check of the first surgery before the second operation and (ii) a gap of more than 14 days between the first operation and the last postoperative check for the first operation | Process | 87.63 | 23.21 |
Bladder tumor |
i01-02 | Vesicoclysis: percentage of non-muscle invasive bladder cancer patients with a trans-urethral resection of the tumor (TURT) that have a washing out of the urinary bladder within 24 h after the TURT | Structure | 69.98 | 23.66 |
Adenoid and tonsils |
i10-02 | Preoperative consultation: percentage of tonsillectomy patients screened at an anesthesiology outpatient clinic before tonsillectomy | Process | 90.80 | 23.58 |
i10-04a | Postoperative pain measurement: percentage of inpatient tonsillectomy patients that have their pain intensity measured every 8 h | Process | 84.76 | 21.61 |
i10-04b | Postoperative pain measurement: percentage of measured inpatient patients without serious pain (i.e., Visual Analog Scale for Pain \(\le\) 7 or Numeric Rating Scale \(\le\) 7) | Process | 93.32 | 14.51 |
i10-04c | Postoperative pain measurement: percentage of daycare patients that have been telephoned after their operation to monitor their pain intensity | Process | 75.47 | 38.63 |
The indicators measure various aspects of the quality of care. For example, for cataract i02-02 measure the complication rate, while i02-03b measure the diagnosis process. We can indeed observe that (i) i02-02 have a higher average score than i02-03b and (ii) i02-02 have a lower standard deviation than i02-03b. This is unsurprising, since a one percentage-point change in i02-02 has more direct clinical relevance than a one percentage-point change in i02-03b. For this reason, it is more useful to compare changes in terms of one standard deviation. Note that i02-02 is an outlier with respect to the standard deviation. The other indicators have more similar standard deviations and clinical relevance.
In order to compare the quality indicators over the years and with one another, we rescaled and aggregated the indicators as follows: for each year
t \((t=1,2,\ldots ,n)\) and diagnosis group
k (
\(k=1,2,\ldots ,K\)) we calculated a combined quality indicator score per hospital
i (
\(i=1,2,\ldots ,N\)). First, each indicator
h \((h=1,2,\ldots ,H)\) of, say, diagnosis group
k is rescaled to a
z-score (
z-score of indicators have an average value of zero and a standard deviation of 1):
$$\begin{aligned} z_{i{th}}= \frac{(P_{i_{th}}- \mu _{ th})}{sd_{ th}} \end{aligned}$$
where
\(z_{ith}\) is the value of the indicator
h of diagnosis group
k for hospital
i in year
t,
\(\mu_{th}\) is the average value of indicator
h in year
t and
\(sd_{th}\) is the standard deviation of indicator
h in year
t. Note that for all indicators that a high
z-score is associated with high quality of care and low
z-score is associated with low quality of care.
Secondly, we calculate for each hospital in each year we calculated an average
z-scores for diagnosis group
k, which we denote by
\(\mathrm{quality}_{itk}\), by averaging over
\(z_{ith}\):
$$\begin{aligned} \mathrm{quality}_{itk}= \frac{\sum _{h=1}^{H} z_{ith}}{H}. \end{aligned}$$
We thus interpreted
\(\mathrm{quality}_{itk}\) as the diagnosis group
k (combined) quality indicator of hospital
i in year
t.
Data
This research was based on claims data from 2008 to 2011 from all Dutch hospitals and Independent Treatment Centers (ITCs). The Dutch hospital market consists of 87 hospitals, two specialized hospitals, and eight academic hospitals. The total number of ITCs rose from 189 in 2008 to 282 in 2012. Our unique dataset consists of patient-level data including patient characteristics such as gender, age, zip code, diagnosis and treatment, hospital characteristics and hospital contract prices for our three selected diagnoses. The total number of patients in the period 2008 and 2011 for cataract was 474,410 with a total revenue of 843 million euros. The total number of patients in the period 2008 and 2011 for adenoid and tonsils was 223,177 with a total revenue of 219 million euros. The total number of patients in the period 2008 and 2011 for bladder tumor was 46,497 with a total revenue of 244 million euros.
The calculation of the weighted market share was based on the claims data from the hospitals and Independent Treatment Centers (ITCs), where we removed those cases with invalid zip codes (which amounted to less than 1% of our sample). Thus, when calculating the market share, we were able to take into account the (potential for) competitive pressure from ITCs. However, as discussed in “
Quality indicators”, we had no data on the quality indicators for the ITCs. This implies that the analysis of the quality indicators is restricted to hospitals and excludes the ITCs. For each diagnosis, we removed those hospitals where there was no data on quality indicators. Tables
2,
3, and
4 present the descriptive statistics for the variables used in our empirical model at the diagnosis level. The total number of observations ranges from 191 for tonsils to 286 for cataract. For example, the average actual market share
ms for cataract was 0.58 (SD 0.21) with a minimum of 0.06 and maximum of 0.97 meaning that there are large differences between the market share of hospitals. The other diagnoses display comparable actual market share. The average hospital-insurer HHI is moderately strong and ranges from 0.39 for cataract, 0.32 for tonsils, and 0.38 for bladder tumor.
As mentioned in “
Quality indicators”, we have used standardized quality scores for each diagnosis. The standardized quality scores ranged between
\(-3.05\) and 0.78 for cataract, between
\(-2.17\) and 0.71 for tonsils and between
\(-2.85\) and 1.42 for bladder tumor.
We included two hospital characteristics in our empirical model. Firstly, we used a dummy variable to control for whether a hospital is a general hospital or academic (university) hospital, a low-volume dummy, and insurance-hospital HHI. For each diagnosis, less than 10% of the hospitals included were academic hospitals. To control for patient characteristics, we included three variables: the fraction of female patients, the fraction of patients that were older than 65 years, and the co-morbidity index. The results differed per diagnosis due to the characteristics of the condition and the set of hospitals included in the analysis. The co-morbidity index, which is defined in this study as the average number of diagnoses, varied considerably between the diagnoses. The co-morbidity for tonsils had the lowest index of 1.04. The average number of additional diagnoses for bladder tumor (3.19) and cataracts (2.18) was higher due to the older population involved.
ms
| 286 | 0.58 | 0.21 | 0.06 | 0.97 |
Quality | 286 | 0.00 | 0.63 | −3.05 | 0.78 |
frac_female | 286 | 0.59 | 0.03 | 0.51 | 0.68 |
frac_65 | 286 | 0.83 | 0.06 | 0.58 | 0.91 |
Com | 286 | 2.18 | 0.30 | 1.60 | 3.17 |
HHI_ins | 286 | 0.39 | 0.11 | 0.19 | 0.63 |
Lowvolume | 286 | 0.25 | 0.43 | 0 | 1 |
Acad | 286 | 0.09 | 0.28 | 0 | 1 |
Table 3Adenoid and tonsils
ms
| 191 | 0.69 | 0.19 | 0.12 | 0.98 |
Quality | 191 | 0.00 | 0.59 | −2.17 | 0.71 |
frac_female | 191 | 0.51 | 0.03 | 0.42 | 0.60 |
frac_65 | 191 | 0.003 | 0.003 | 0.00 | 0.02 |
Com | 191 | 1.04 | 0.30 | 0.53 | 2.18 |
HHI_ins | 191 | 0.32 | 0.09 | 0.17 | 0.57 |
Lowvolume | 191 | 0.25 | 0.43 | 0 | 1 |
Acad | 191 | 0.07 | 0.26 | 0 | 1 |
ms
| 199 | 0.73 | 0.15 | 0.33 | 0.98 |
Quality | 199 | 0.00 | 1.00 | −2.85 | 1.42 |
frac_female | 199 | 0.22 | 0.05 | 0.10 | 0.45 |
frac_65 | 199 | 0.72 | 0.06 | 0.52 | 0.86 |
Com | 199 | 2.68 | 0.46 | 1.70 | 4.32 |
HHI_ins | 199 | 0.38 | 0.11 | 0.19 | 0.71 |
Lowvolume | 199 | 0.23 | 0.42 | 0 | 1 |
Acad | 199 | 0.10 | 0.29 | 0 | 1 |