A.  Patient Cost Sharing and Provider Reimbursement

Taiwan’s National Health Insurance (NHI) database is ideal for this study for several reasons. First, similar to Canada, the Taiwanese NHI is a single-payer system for all citizens and residents consisting of a uniform, comprehensive care benefits package that covers drugs, hospitals, and primary care (Chen and Chuang 2016). Since participation in the NHI is mandatory, we can eliminate doubts about adverse selection issues. Furthermore, we can address patient selection issues because the database includes beneficiaries who have never been in the hospital.

The NHI administration manages health expenditure inflation by reimbursing providers rather than charging deductibles or capping out-of-pocket expenses. This form of reimbursement is fee-for-service via a uniform national fee schedule to prevent providers from selecting patients or imposing price discrimination on them. Since hospitals pay doctors on a fee-for-service basis plus a basic salary that varies across hospitals, the financial incentives of doctors and hospitals are similar.

Moreover, the NHI system imposes a minor penalty (only US$7 in 2014) on hospital visits without primary care referrals. Consequently, almost all patients choose their attending doctor without a primary care referral. As patients can freely check into different hospitals or request different doctors in the same hospital, we analyze doctor-patient relationships by examining hospital admissions data. Taiwanese hospitals work on a closed-staff structure, with the on-staff doctor assuming full responsibility for patient care. This institutional setting ensures that matching patients to physicians can precisely describe the interaction between doctors and patients during hospital admissions.

B.  Data Linkage

We merge several administrative data sources in the NHI database from 2000 to 2018 using unique scrambled identifiers (IDs) via four steps. First, we link the Cancer Registry to the Death Registry and the Registry of Beneficiaries. This data link covers cancer patients’ diagnosis date(s), cancer site(s), and diagnosis stage. It also documents treatment methods, demographic background (sex, birthday, income bracket, and registration district), death records if patients died by the end of 2019, and whether they received hospital care.

Second, we identify the physician-patients and obtain their medical specialties by merging the data with the Registry for Medical Personnel and Records of Board-Certified Specialists using their IDs. The former covers physician gender, birthday, and certification date, and the latter records their medical specialties and practice locations over time. Third, we combine the above data with the Reimbursement Claim Records to obtain inpatient care details per hospital admission 1 year after a cancer diagnosis. These data reveal the entire history of treatments, care volumes, hospital type and location, hospital ID, and attending doctor’s ID so that we can calculate the total inpatient care costs, coinsurance payments, and spending on medicines, surgery, tube feeding, radiation therapy, and examination to construct covariates and outcome variables. Finally, we again use the attending doctor’s ID to derive his or her certified specialty and experience by linking the compiled data to the Registry for Medical Personnel and Records of Board-Certified Specialists.

C.  Time-Varying Doctor Selectivity

Like physicians’ experience, doctor selectivity can vary over time. We estimate an expert patient’s knowledge of doctor selectivity at the time of diagnosis using the percentage of physician-patient hospital admissions 3 years before. For instance, if a doctor attended 1,000 hospital admissions in the past 3 years and only two were physician-patients, the selectivity measure takes a value of 0.002. While the public knows about physicians’ experience, doctor selectivity is typically not well known except to expert patients.

Doctors chosen by physician-patients are twice as selective as those chosen by other patients, with a selectivity level of 0.0039 ($0.0022+0.0017$), as shown in table 1. Moreover, physician-patients select more experienced doctors than nonphysician-patients by 2 years. These are large and statistically significant differences at the 95% level.

A significant challenge of our empirical work is that doctors could become more selective as they gain experience. As a result, patients treated earlier are not necessarily comparable to those treated later by the same doctor. To remove this time-varying bias, we fix the attending doctor and admission time for a fair comparison.

D.  Descriptive Statistics

Table A1 compares the characteristics of physician-patients to those of nonphysician-patients, including their demographics, inpatient care receipt, and survival outcomes. The data comprise more than 1.2 million cancer diagnoses, approximately 1 million patients, and 1,987 medical doctors. The number of cancer diagnoses exceeds that of cancer patients because a single patient can be diagnosed more than once for recurrence or confirmation. Only 0.01% of diagnoses involve multiple cancers. Thirty percent of all cancers first diagnosed from January 2004 to December 2016 were in the advanced stage.³ The data period began in January 2004, when Taiwan began to adopt the AJCC Cancer Staging Manual, the benchmark for classifying cancer patients. Our analysis covers all of the cancer sites listed in table 2.

Notably, statistics indicate that 12% of patients diagnosed with cancer need no hospital care. About a quarter of these are in the advanced stage (not shown in the table). After controlling the interaction of patient demographics, prior medical spending, and admission year, we find that physician-patients are significantly more likely to receive hospital care by 1 percentage point (with a standard error of 0.006; not reported in the table). This difference was statistically insignificant and decreased when limiting the sample to advanced-stage cancer at the first diagnosis.

Each cancer diagnosis was likely to lead to more cancer therapies, including surgery, chemotherapy, radiation therapy, palliative care, targeted therapy, hormone therapy, immunotherapy, stem cell treatment, and Chinese medicine. We excluded the last three from our main analysis because less than 1% of diagnoses led to their adoption (table A1).

Because the Death Registry for our benchmark study is available until December 2016, the N-year survival indicator needs to forgo N years of the combined data. After the first diagnosis, more than 80% of cancer patients survive beyond 180 days, and about 60% live more than 3 years.

One concern about the data is that doctors may have diagnosed physician-patients with advanced cancer earlier than other patients, which would result in overstating physician-patients’ treatment and survival advantages. However, there is no evidence of this in the statistics (table A1). Our analysis includes all of the hospital admissions of first-diagnosis advanced cancer patients. The first diagnosis of physician-patients is about 3 percentage points more likely to be advanced than other cancer patients. This drops below 0.7 percentage points (with a standard error of 0.009 clustered at patient levels; not shown in the tables) after holding constant the patient’s sex, age, income, region, spending on inpatient care, and diagnosis year. These results suggest that our data are unlikely to contain potential bias due to earlier diagnoses by physician-patients.

Hospital admissions of physician- and nonphysician-patients with advanced cancer are compared in table 1, with standard errors clustered at patient levels. These data cover 1,123,377 admissions associated with 279,399 nonphysician-patients and 2,454 associated with 611 physician-patients. Given the closed-staff structure of Taiwanese hospitals, each admission matches one attending doctor to one patient. Statistics indicate that physician-patients are older and wealthier, tend to be male, and have spent less on hospital care before the first cancer diagnosis. Both physician- and nonphysician-patients are equally likely to have a preexisting clinical relationship with a doctor, but physician-patients tend to opt for male or more experienced doctors practicing in single locations and specializing in a cancer-related area or working in a cancer-related department.

Nonphysician-patients wait an average of 122 days to receive inpatient treatment after the first diagnosis, which is 5.59 days longer than physician-patients. This difference is significant at the 90% level. Nonphysician-patients also stay in acute inpatient care units for about 7.89 days, while physician-patient stays are 10% (0.81 days) shorter at the 95% significance level.

The unconditional mean difference tests in table 1 show that physician-patients are less likely to undergo surgery and chemotherapy by 8% and 5% (0.05/0.66; 0.04/0.8) but considerably more likely to use targeted treatment by 44% (0.05/0.11). However, these gaps may be due to different health or socioeconomic conditions or the selection of doctors with different practice styles.

Finally, the bottom part of table 1 shows that physician-patients have the same 180-day survival rate as other patients. However, their 1- and 3-year survival rates are higher, possibly due to higher income, better health, better communication, closer relationships with attending doctors, selection of doctors, or more cancer-related knowledge. The higher survival rate of mainly older male physician-patients than other patients with advanced cancer seems inconsistent.

A.  Balance Checks

First, the attending doctor is left unmatched, and nonphysician-patients and physician-patients of the same patient type in the same hospital are compared. Table 3 shows the balance checks for two matching schemes: scheme A (left) considers the exact match for patient kinds within hospitals, and scheme B (right) looks at those within doctor-hospital. Scheme A excludes 98% of nonphysician-patients and 84% of physician-patients because of nonoverlap in the covariate cells. Overlap is extremely rare in matching physician-patients with other patients with advanced cancer because physician-patients are significantly older, healthier, wealthier, and more masculine. After matching, the total number of admissions is 2,811, comprising 685 admissions (for 98 matched physician-patients) versus 2,126 admissions (for 565 matched nonphysician-patients).

Although scheme A narrows down comparable patients, these patients are mostly attended by different doctors; hence, the varied outcomes observed between physician-patients and other patients may be due to physician quality. We improve the balance of matches by further matching attending doctors in scheme B, which reduces the sample size to 552 admissions. Of these admissions, 252 are for 31 physician-patients and 300 are for 69 nonphysician-patients.

The balance tests for variables we did not match are shown in table 3, in which the p-values of the mean difference (t-tests) and distributional difference (Kolmogorov-Smirnov [KS] tests) are reported. Both tests have p-values equal to 1 for the precisely matched covariates. Scheme A statistics indicate that patients’ prediagnosis health conditions, estimated by pretrends in inpatient costs and prior drug spending, are balanced statistically. In contrast, doctors who treated physician-patients have 0.3 standard deviations more experience than those who treated nonphysician-patients. The distribution of the doctor’s gender, mobility, and specialties differs significantly between physician- and nonphysician-patients.

After matching patients according to their attending doctors in scheme B, none of these prediagnosis attributes indicate significant gaps in sample means or distributions. This result strongly suggests that matching attending doctors substantially improves the balance of observables, making it plausible that unobserved confounders also balance out.

B.  Matching Estimates

Table 4 reports the matching estimates for the two matching schemes: within-hospital comparison between 2,811 matched admissions (scheme A) and within-doctor-hospital comparison between 552 matched entries (scheme B). The standard deviation in outcomes is displayed in columns 1 and 5 after removing the variation of the matched covariates. Further matching those 2,811 admissions with attending doctors in scheme B reduces the standard deviation by 15%–75%. This indicates that many of the changes in outcome are due to the variation in attending doctors.

As scheme A does not match the hospital entries according to attending doctors, physician-patients tend to see more experienced and selective doctors than their nonphysician counterparts in this scheme (table 3). Suppose that physician-patients prefer fewer tests and fewer intensive therapies at the advanced cancer stage and that experienced or highly qualified doctors tend to choose more intensive types of care and order more tests.⁶ As physicians can more easily identify highly skilled doctors than nonphysicians, we understate physician-patients’ negative impact on intensive care utilization and checkup costs if we do not match them according to their attending doctors.

Matching hospital entries according to attending doctors and hospitals in scheme B dramatically increases physician-patients’ surgical/radiation adoption and costs for examinations, as expected. Physician-patients are 8 percentage points less likely to undergo surgery and 7 percentage points less likely to choose radiation therapy. These statistically significant estimates account for 42% and 21% of the residual standard deviation (0.083/0.20; 0.071/0.33). In contrast, scheme B reduces the margins of intensive care. Physician-patients’ tube-feeding care drops from approximately 0.3 to 0.03 percentage points. The effect of radiation also substantially decreases and becomes statistically insignificant. These impacts of different treatments on physician-patients’ survival within-hospital and within-doctor-hospital matched samples suggest that physician-patients pick better doctors, even within hospitals.

Our benchmark (scheme B) shows that physician-patients are significantly less likely to adopt surgery by 0.4 standard deviations (0.083/0.20) and radiation therapy by 0.2 standard deviations (0.071/0.33). As for intensive margins, physician-patients utilize surgery less frequently than their counterparts by 0.4 standard deviations (1.159/2.87) while taking approximately the same radiation dose. In addition, while using less intensive care, physician-patients with advanced cancer are less likely to adopt palliative care by 0.2 standard deviations (0.027/0.16). The only items physician-patients utilize more are targeted drug therapy ($0.167/0.28=60\text{\%}$ more likely) and prescription medication ($0.652/1.80=0.4$ standard deviations greater costs).

Physician-patients with advanced cancer spend more on medication, but the source of the spending rise cannot be determined from the currently available data. The higher spending could be due to greater quantities, more varieties, or increased drug prices (e.g., on patented brands). Doctors/hospitals cannot charge different prices for the same drugs because of universally uniform reimbursement prices and adjustments (Chen and Chuang 2016). This institutional feature leaves physician-patients’ increased drug dose or variety as the likely explanation for their increased medication cost.⁷

C.  Cost-Effectiveness

According to the American Cancer Association’s published medical guidelines, surgery/radiation therapies are more appropriate for early-stage cancers. More advanced cancer requires treatment that reaches the entire body (e.g., chemotherapy and targeted drug therapy). We have shown that physician-patients receive less surgery/radiation treatment for advanced cancer than the matched nonphysician-patients while spending more on drugs and being more likely to use targeted therapy. If physician-patients’ treatment is clinically suitable, our results indicate that underuse and overuse coexist among nonphysician-patients.

Physician-patients receive different and better treatment, significantly increasing their chances of short-term survival compared with other patients by 2.5 (9.3) percentage points at 180- and 365-day thresholds, as shown in columns 6 and 7 of table 4. Their long-term survival is also higher by 7.1 percentage points at the 3-year cutoff. These estimates account for at least 0.25 standard deviations. Besides the survival benefit of better treatment, physician-patients pay significantly less for coinsurance by 0.226 log points. Overall, physician-patients receive cost-effective care relative to that given to matched patients.

A.  Physician-Patients Are Diagnosed Earlier or Treated Earlier

Physician-patient relationships and informational advantages may lead to earlier diagnoses or treatment than for nonphysician-patients, resulting in the former needing less intensive care and surviving longer. However, the hypothesis that the physician-patient status reduces the likelihood of being diagnosed too late cannot be accepted, as discussed in section II.D.

The matching estimates for scheme B in table 4 show that physician-patients wait 1.3 days longer than other patients between diagnosis and treatment. This difference is statistically insignificant and accounts for less than 2% (1.3/75.6) of a standard deviation. Hence, the hypothesis that physician-patients receive treatment earlier than nonphysician-patients cannot be accepted.

B.  Physician-Patients Exhibit Better Health

To ensure that the matched physicians and other patients have similar health conditions, we compare only patients in the same quintile for hospital spending in the 3 years before being diagnosed with advanced cancer. However, it remains possible that physician-patients are healthier than their counterparts in a way not captured by our model. We test this hypothesis by checking the balance of variables excluded from the covariate list. The placebo test results in table 3 are contrary to this hypothesis. The matched patients’ previous drug spending and pretrend hospital costs are not significantly different. This pattern is robust, irrespective of scheme A or B (accounting for the attending doctor or not) if limited to matched patients in the same hospitals who share the same attributes and cancer sites during the same admission period. This result suggests that physician-patients’ better health is not due to reduced surgical/radiation treatment before diagnosis.

C.  Physician-Patients Are Less Likely to Sue for Malpractice than Other Patients

Another explanation for our finding of physician-patients’ reduced intensive care and examinations is that they are less likely than nonphysician-patients to sue for malpractice. Currie and MacLeod (2008) suggest that potential liability concerns may induce doctors to perform more unnecessary procedures, especially for nonphysician-patients in our context. We examine this explanation by investigating the frequency of malpractice lawsuits for our matched data.

During our data period (2004–16), medical doctors in Taiwan were subject to no-fault or joint and several liability (Ministry of Health and Welfare 2018). ER doctors, neurosurgeons, and anesthesiologists were most likely to be sued and pay noneconomic damages (Chen et al. 2012). On investigating our matched data to see whether physician-patients in the riskiest departments receive exceptionally high premiums, we found almost no physician-patients with cancer seeking care in these departments. As a result, there is no evidence that the lower utilization rates of surgical or radiation therapy of physician-patients with advanced cancer could be explained by defensive medicine. Nevertheless, fear of litigation may still drive some doctors to prescribe different types of procedures to physician-patients than nonphysician-patients due to unobserved differences, which we will address next.

A.  Relational Mechanism: Mapping Specialists to Social Networks

As the attending doctor is fully responsible for caring for each admitted patient under Taiwanese NHI’s closed staff structure, each doctor-patient pair has a well-defined professional connection. Physician-patients who share the doctor’s specialty have a stronger tie because they may have met professionally before the diagnosis. Since subspecialties are unobservable in our data, we can infer patient-doctor networks from physician specialties, TOS taxonomy classification, and the doctor’s hospital department. We follow the TOS taxonomy to classify networks, allowing one doctor to have multiple networks, and summarize how we map specialists according to a social network in table 2.

TOS covers eight of the top 10 cancer treatment specialties: three from the medical oncology network (category I, cols. 1–3) and six from the surgical oncology network (category II, cols. 3–8), with radiation oncology belonging to both networks (col. 3).¹¹ Although TOS lists pathologists as part of the medical oncology network, we separate from the medical oncology network those without another specialty and those who have never worked in internal medicine departments. However, despite having no medical knowledge of or caseload involving cancer treatment, TOS-listed plastic and reconstructive surgeons are included as part of the surgical oncology network because of the double-dose surgical training required for them to become board-certified surgical oncologists. Finally, although dermatologists and ophthalmologists who treat most skin/eye cancers are excluded from the TOS taxonomy (cols. 9, 10), we assume that they have separate networks unrelated to cancer care in areas denoted by “others.”

B.  Informational Mechanism: Mapping Specialists to Medical Knowledge of a Cancer Site

We estimate physician-patients’ medical knowledge of their cancer site using the caseload per doctor in the same specialty relative to the highest volume. Table A3 illustrates our estimation process for advanced breast cancer. First, we calculate the caseload per doctor by specialty and hospital department using 100% inpatient reimbursement data (col. 1); then we measure the relative knowledge/caseload of the cancer site per specialty across hospital departments (col. 2). The top breast cancer experts are surgeons in surgery departments with the highest caseload, while doctors in other departments/specialties account for only a fraction of the top expert caseload.

Physician-patients’ previous workplaces or hospital departments are unobservable. We then estimate their knowledge of the cancer site given their first specialty in column 3 by averaging the relative caseloads per specialty across departments, weighted by department shares. The mapping of each cancer site to specialty-specific knowledge is summarized in table 2. For example, the top advanced breast cancer experts are surgeons in surgery departments whose knowledge level of this cancer is 0.49, far below 1, since most surgeons working in departments rarely treat breast cancer. This data feature is common among other specialists (e.g., radiation oncologists).

We apply this method to all cancer sites. Surgeons with specialties not listed in the columns have zero relative knowledge/caseload (e.g., neurologists, neurosurgeons, and plastic and reconstructive surgeons). The statistics show that the top cancer experts are typically organ-specific specialists (except for bone cancer), and radiation oncologists are the top leukemia and stomach cancer treatment experts. Finally, we standardize physician-patients’ knowledge levels of their cancer site.

C.  Exploration of Mechanisms

Using the network indicator R and continuous information measure K constructed in sections V.A and V.B, we evaluate the relative contributions of relational and informational advantages to cancer treatment and patient survival. Let $\beta (R,K)$ denote physician-patients’ heterogeneous impact on outcomes given their network R and knowledge K. We assume that the full effect can be broken into four components:

To evaluate these mechanisms, we first identify the physician status effect (β) by comparing nonphysician-patients and physician-patients with no advantage. Second, we estimate the total impact of a professional tie with the attending doctor ($\beta (\mathrm{1},K)-\beta (\mathrm{0},K)$), including the network’s main effect and the network-induced informational effect ($\rho +\mathrm{E}[\delta (K)]$), so relational and informational mechanisms are both at work. We cannot separate these two effects using matching methods due to the lack of support of continuous medical knowledge.¹² We overcome this challenge in the third step by estimating the value of a network by information percentile when applying fixed effects models to matched physician-patients. Each step is expanded, as shown below.

3.  Interchangeability between Network and Knowledge

Although the average physician-patient effect on some treatments is consistent with the impact of physician status or network-related mechanisms, neither of these elements explains why average physician-patients tend to replace radiation with targeted therapies while enjoying better short-term survival than nonphysician-patients. This suggests that this information is crucial in determining treatment and survival. Using equation (1), we quantify the value of a network by information percentiles using the degree of informational interchangeability for a network, that is, how many more medical knowledge percentiles are required to maintain the same survival or treatment intensity if the physician-patient lacks a robust professional tie with the attending doctor. Our parameter of interest is

$$\begin{array}{}\text{(2)}& V\left(R,K\right)=-{\frac{dK}{\mathrm{\Delta }R}|}_{\mathrm{fixing}{\beta }_{KR}}=\frac{\beta \left(1,K\right)-\beta \left(0,K\right)}{\partial \beta \left(R,K\right)/\partial K}=\frac{\rho +\delta \left(K\right)}{e^{\prime }\left(K\right)+R{\delta }^{\prime }\left(K\right)}\mathrm{.}\end{array}$$

The value of networking can vary with physician-patients’ professional connections ($R=0$ or 1) and knowledge levels ($K\in [0,1]$). We note that $V(1,K)$ equals $V(0,K)$ at a knowledge threshold of $K=k$, where ${\delta }^{\prime }(k)=0$.¹⁶ Deviating from the threshold, δ(K) could be downward or upward sloping or nonlinear in K. To allow flexible nonlinearity, we assume polynomials with orders $(4,4)$ for δ(K) and e(K). If using orders $(4,3)$ or $(3,4)$, we find little shift in the resulting pattern.

We perform fixed effects regressions with two causal variables, network and knowledge, by combining two balanced samples to derive sufficient variation: one balancing patients with versus without a network (as shown in cols. 1–3 of table A5), and the other balancing knowledge above versus below the median (cols. 4–6). We illustrate the data structure in table 6. The data balanced by networks has 277 observations, which overlaps the data balanced by knowledge in 214 cases. The data balanced by knowledge adds 80 extra admissions that are not balanced by network(s).

All Physician-Patients	All Physician-Patients		Total
	Not Matched	Matched Subsample for Patients With versus Without a Network
Not matched	2,656	63	2,719
Matched subsample for patients with knowledge above vs. below the 50th percentile	80	214	294
Total	2,736	277	3,013

View Table Image

Using this integrated sample, we estimate that the fixed effects model holds the same list of covariates constant as in scheme C of section V.C.1. Although the estimated coefficients are imprecise (tables A6-1, A6-2) due to the limited support of continuous information, there are clear patterns in the estimated value of networks. We sketch the interchangeability of a given network $V(1,K)$ in figures 1–3, as the case with no network, $V(0,K)$, is too noisy.

We find a strikingly persistent pattern. Lacking relevant knowledge of the cancer site makes networking equivalent to knowledge reduction. The value of networking for most outcomes, when the knowledge level is at or below the bottom quartile, is significantly negative, between −11 and −6 percentage points. It further expands to −18 and −14 percentage points at the bottom 1st percentile. For professionally connected but less informed patients, the social tie is equivalent to losing 15–18 percentage points of relevant knowledge of the cancer site to maintain the same survival or treatment intensity. This result affects survival rates, examination costs (fig. 1), use of surgery/radiation therapy at extensive and intensive margins (fig. 2), adoption of target drug therapy, and drug/tube-feeding costs (fig. 3). Palliative care and chemotherapy are exceptions for which we obtained insufficient precision.

In summary, when doctors attend to professionally connected physician-patients with no relevant knowledge, they may abuse the trust of these patients by deviating from the treatment they would prescribe to a knowledgeable specialist-patient, harming their chance of survival.