Abstract
It is well known that the differences-in-differences (DD) estimator is based on the assumption that in the absence of treatment, the average outcomes for the treated group and the control group will follow a common trend over time. That can be problematic, especially when the selection for the treatment is influenced by the individual’s unobserved behavior correlating with the medical utilization. The aim of this study was to develop an index for controlling a patient’s unobserved heterogeneous response to reform, in order to improve the comparability of treatment assignment. This study showed that a DD estimator of the reform effects can be decomposed into effects induced by moral hazard and by changes in health risk within the same treated/untreated group. This article also presented evidence that the constructed index of the price elasticity of the adjusted clinical group has good statistical properties for identifying the impact of reform.
Keywords
In recent years, a number of studies have used a natural experimental design with a differences-in-differences (DD) estimator to examine the impact of copayment reform (Chiappori, Durand, & Geoffard, 1998; Cockx & Brasseur, 2003; Rosen, Brammli-Greenberg, Gross, & Feldman, 2010; Trivedi, Moloo, & Mor, 2010; Winkelmann, 2004a). Under this approach, the crucial assumption is that the average outcomes for the treated group and the control group will follow a common or parallel trend (Abadie, 2005; Jones, 2009; Jones & Rice, 2009). In other words, the DD estimator requires that in the absence of reform, the changes in the treatment group are equal to the observed changes in the control group, so that by comparing changes the observed and unobserved time-invariant individual characteristics that may be correlated with medical expenses can be controlled. However, this poses the risk that the pretreatment characteristics that are thought to be associated with the dynamics of the outcome variable may not be balanced between the treated and the untreated groups. For example, if the selection for treatment is influenced by the individual’s socioeconomic status which is correlated with medical utilization, then that will cause a different outcome trend between the treated and the control groups (Jones & Rice, 2009; Riphahn, Wambach, & Million, 2003). Thus, it is not surprising that recent studies appear to have mixed results regarding the effect of copayment (Chiappori et al., 1998; Cockx & Brasseur, 2003; Rosen et al., 2010).
Rather than constructing a comparison group designated by legislation and thus being subject to the major criticism of lacking random assignments, we proposed an alternative that serves a dual purpose. First, it identifies an individual’s price elasticity regarding genuine need for health care so that it can complement an existing natural control group (NCG) and enhance the comparability between treatment and control group by utilizing the differences-in-differences-in-differences (DDD) approach. Second, as will be demonstrated later in this article, it passes the placebo test for the parallel trend assumption that underlies the DD approach. It can thus serve as the ideal candidate for treatment assignment when there is no NCG. Specifically, we developed an index, employing the Johns Hopkins Adjusted Clinical Group (ACG) Case-mix System for controlling the patient’s unobserved heterogeneous response to reform. We defined the comparison group as that group of patients whose ACG price elasticity (ACGPE) is <1, in order to capture the fact that their health care utilization may be more strongly driven by genuine health concerns and less by financial incentives. There is a substantial body of literature indicating that the ACG system is well capable of predicting health care needs (Chang, Lee, & Weiner, 2010; Lemke, Weiner, & Clark, 2012; Reid, Roos, MacWilliam, Frohlich, & Black, 2002). Nevertheless, to the best of our knowledge there is not yet an application that aims to improve the comparability of treatment assignment within a natural experimental framework.
Method
Data Sources
Data were taken from the Longitudinal Health Insurance Database provided by the National Health Research Institute (NHRI) in Taiwan for the period January 1, 1998, to December 31, 2000. This period covers a major change in policy in August 1, 1999, with respect to the copayment required for ambulatory care and prescription drugs. To avoid any transition effects, we dropped the 1999 observations from the sample. Consequently, at least 1 year of enrollment before (1998) and after the reform (2000) had to be included in this analysis for the purpose of comparison. It should be noted that although the dates used (1998 and 2000) are old, they have no relevance to the purposes of this study. In addition, two different sample sets from the NHRI were employed to validate that the treatment assignment satisfied the parallel trend assumption underlying the DD approach. Each sample contained enrollment and claims files from a randomly chosen 0.2% (referred to as Sample A) and 0.15% (Sample B) of Taiwan’s total population (approximately 23 million individuals).
The enrollment files contain the individual subscription information and demographics, including sex, date of birth, and the category of the beneficiary based mainly on occupation and income. The claim files contain comprehensive records of the utilization of ambulatory care, including date of service, the International Classification of Diseases, Ninth Revision, Clinical Modification diagnosis codes, claimed medical expenses, and the amount of copayment for each medical visit. Observations with missing values for any of the dependent or independent variables were deleted. The final sample sizes were 123,532 individuals for Sample A and 95,880 for Sample B.
Treatment Assignment
Control group from a natural experiment: Exempt versus nonexempt. Individuals who are explicitly exempted from the increase in copayment include those with a low income as qualified by the Social Support Law, as well as veterans and their survivors. This natural experiment allowed us to employ these people who were unaffected as the control group (referred to as an NCG) for DD estimation.
However, the division into a treatment and control group based on such natural experiment may be problematic (Winkelmann, 2004a; 2004b). In a National Health Insurance (NHI) scheme, such as the one in the present study, the exemption status in most cases is designated by that individual’s socioeconomic status, which in turn is responsible for the degree of health care utilization. For example, income is likely to be subject to considerable selection bias and reporting error because low-income earners may have a poorer health status, resulting in a higher rate of increased outpatient visits in this NCG (Jones, 2009; Jones & Rice, 2009; Rosen et al., 2010).
Counterfactual construction by the ACG morbidity measure
Besides the NCG described earlier, we constructed an index of the ACGPE for health care to define a comparison group that is less likely to be affected by copayment.
First, we employed the ACG Case-mix System to assign each patient an ACG code reflecting their health needs. Second, we calculated the level of resources necessary for delivering appropriate health care to an ACG as the total ambulatory expenses divided by the total number of visits for all patients with the same ACG. This then is the average price per visit of an ACG (AVPACG), that is,
As individuals with the same ACG are similar in terms of both clinical criteria and expected need for health care resources, this AVPACG is assumed to be stable across time.
Third, we computed each individual’s ACGPE for doctor visits as the increasing (or decreasing) rate of their visits divided by that of the AVPACG for 2 years prior to the reform, that is,
An ACG index not only captures the specific clustering of morbidities experienced by a person over a given period of time, it also predicts future health care utilization assumed to be correlated with the illness burden of that individual. Therefore, the ACGPE for doctor visits can be used for a more equitable comparison of utilization or outcome across two or more patients. Although two patients may have the same ACG code because of their similar illness burden and expected health needs, their number of doctor visits may be substantially different because they may differ in their behavior of seeking health care. Thus, patients with the same ACG may have a different ACGPE for doctor visits. An individual with an ACGPE <1 indicates that his or her health care utilization is relatively insensitive to the AVPACG. This in turn suggests that an individual’s number of doctor visits goes hand-in-hand with the medical cost required for delivering proper health care to an ACG, which is a proxy for an individual’s genuine health needs. We thus defined patients with an ACGPE <1 as the ACG control in order to capture the fact that their health care utilization may be more strongly driven by genuine health concerns and less influenced by financial incentives. On the other hand, an individual with an ACGPE >1 implies that other than his or her health status or expected health needs is subject to some unobserved behaviors, such as moral hazard, which can cause a greater response to the AVPACG. Individuals with an ACGPE >1 are defined as the ACG treatment group.
Dependent Variable
The hope was that the copayment reform would contain the total outpatient expenditures from the demand side (Manning, Newhouse, & Duan, 1987; Zweifel & Manning, 2000). The literature on the demand for medical care analyzes either discrete measures, such as the number of doctor visits (Chiappori et al., 1998; Cockx & Brasseur, 2003; Riphahn et al., 2003; Winkelmann, 2004a; 2004b), or continuous measures, such as expenditures (Hsu, Lin, & Yang, 2008; Kim, Ko, & Yang, 2005; Lin, Hsu, & Takao, 2008; Manning et al., 1987). However, it is well known that these two measures are derived from two different decision-making processes based on a principal–agent model, where the physician (the agent) determines utilization on behalf of the patient (the principal) once initial contact is made (Deb & Trivedi, 2002; Pohlmeier & Ulrich, 1995).
Therefore, when assessing the effects of the reform on the demand for health services, it is useful to distinguish between a direct and an indirect effect (Winkelmann, 2004a; 2004b). The direct effect is a movement up the demand curve, such as the demand for prescription drugs, because increased copayment directly increase a patient’s out-of-pocket expenses for drug purchases (Winkelmann, 2004a; 2004b). However, this direct effect is largely influenced by the physician’s choice of prescription and style of practice. As pharmaceuticals are available against a prescription and only physicians can issue such a prescription, obtaining a drug is generally preceded by a consultation with a doctor. It has been argued that in countries with a national health insurance system, physicians are likely to respond to the initial fall-off in visits and income by taking steps to increase their service intensity or volume, in part through supplier-induced demand (Evans, Barer, & Stoddart, 1995; Rosen et al., 2010; Zweifel, van der Gaag, & Perlman, 1981). This tends to be more prevalent in a fee-for-service setting, such as in the present article. In the case presented here, it is possible for an increase in total expenditures, even if the demand is completely inelastic or if the number of doctor visits goes down. A World Health Organization report (Saltman & Figueras, 1997) contended that since intensity is largely provider initiated, there is not much scope for cost sharing to make an impact on the overall level of spending.
On the other hand, the indirect effect is a potential inward shift of the demand curve for doctor visits due to the increase in copayment (Winkelmann, 2004a, 2004b). There are good reasons to believe that the indirect effect may be quantitatively as important as the direct one. The rationale is that, typically, consumers exercise substantial discretion (Deb & Trivedi, 2002), not only regarding their first visit but also their subsequent visits, irrespective of new appointments, advice, or a referral from their physician. Several responses to a price increase are possible, including influencing the doctor to provide prescriptions for an extended period of time to avoid charges for multiple visits in the future or not seeing a doctor at all. Both these behavioral changes would reduce the number of visits to a doctor. Alternatively, people might still see a doctor to seek advice on preventive care or see a doctor but decide not to buy the drugs the doctor prescribed. In either case, the number of visits tends to be unaffected by the increased copayment, although the effect on total expenditure will be complicated by the supplier-induced demand as discussed earlier.
For all these aforementioned reasons, the following empirical analysis will provide an estimate of the extent to which the individual demand for doctor visits changed due to the reform. Nevertheless, a future evaluation including evidence on the induced demand for pharmaceuticals is clearly needed.
Statistical Model
The basic empirical strategy is to pool the data over the 2 years in question and estimate the effects of the copayment by comparing the expected number of visits before and after the reform. Because of the count nature of the dependent variable, we used the negative binomial (NB) model, where the right-hand side index of Equation 1 enters an exponential function to model the expected number of visits, as follows:
where yit denotes the number of doctor visits of individual i at time t, and treat i indicates whether individual i belongs to the treatment group, in which the two specifications discussed earlier, that is, the natural treatment group (NTG) and the ACGPE >1 are considered. Similarly, reform t is a dummy variable indicating whether the increased copayment is applicable in period t. The interaction between treat i and reform t denotes the observation of a person in the treatment group after the reform. Vector Xit indicates all other control variables, including gender, a second-order polynomial in age, a density ratio of patients to clinics for six NHI-designated regions where each individual is registered, and two indices for health status: the major illness identity (MI-ID) status and the chronic condition count for the total number of expanded diagnosis clusters (EDCs) that an individual has. The EDCs are obtained from the ACG software. A count of EDCs includes the trigger diagnoses that indicate a significant chronic condition with the expected duration and resource requirements. Individuals with fewer and less severe chronic conditions are healthier and have a relatively low current total cost, with no indication of unexpected future pharmacy costs.
In Equation 1, τ measures the reform effect. If it is negative, the demand for doctor visits in the treatment group falls relative to the demand in the control group after the introduction of the increased copayment.
In addition, we estimated the DDD model as follows:
To further investigate how DDD can be decomposed into effects induced by moral hazard and changes in genuine health needs within the same treated/untreated group, we computed the average treatment effect of DDD using Equation 3.
The NCG and its counterpart each have a DD controlled for the ACGPE. The first two brackets compare the pre-post visits in the natural treatment between ACGPE >1 and ACGPE <1, whereas the last two compare those in the NCG. The simple difference (SD) of the pre-post comparison within the same treated/untreated group when ACGPE <1 indicates the change in medical utilization over time and reflects the individual’s genuine health needs. On the other hand, the SD when ACGPE >1 not only includes the variations in medical utilization over time but also the individual’s behavioral response to the policy reform. Subtracting these two SDs yields the DD, which is an estimate for the moral hazard effect within the group.
Results
Table 1 shows the summary statistics of the variables involved in the analysis using Sample A for both the treatment and the control groups, before and after the reform. The majority of observations are for the treatment group, with approximately 85.8% and 88.7% of the entire sample for the NTG and ACG treatment, respectively. It is evident that the average number of doctor visits shows a slight increase, followed by a modest decrease, which clearly coincides with the timing of the reform.
Descriptive Statistics for ACG Control, Sample A.
Note. Standard deviation shown in parentheses. ACG = adjusted clinical group; ACGPE = ACG price elasticity; EDCs = expanded diagnosis clusters; MI-ID = major illness identity.
The NCG in the present study is comprised of two groups that were exempt: one for low-income earners and the other for veterans and their survivors. As income is likely to be subject to considerable selection bias (Jones, 2009; Jones & Rice, 2009), the latter group is likely to be less error prone and thus may be a superior comparison group by itself, rather than including the low-income group. To demonstrate whether this is true or not, we compared the average characteristics of the treatment group to those of the veterans and their survivors on several independent variables to investigate whether these two groups were alike before the reform. As shown in Table 2, the two-sample t test rejected the hypothesis that the mean values of these control variables were equal. This suggests that the preexisting differences between the two groups are substantial, which may not be comparable at baseline. Thus, veterans and their survivors alone could not serve as a better candidate for the comparison group.
Descriptive Statistics for Natural Control, Sample A.
Note. Standard deviation shown in parentheses. EDCs = expanded diagnosis clusters; MI-ID = major illness identity.
** and *** denote the mean differences between the natural treatment group and the low-income earner/veteran and their survivors at significance levels of p < .01 and p < .001, respectively.
Reform Effects Under Different Treatment Assignment
Two sets of panel regression results, one for the natural control and the other for the ACG control, are shown in Table 3. All analyses were performed using STATA version 10.0. The parameters of the fixed-effects negative binomial (FENB) model were estimated using the conditional maximum likelihood function proposed by Hausman, Hall, and Griliches (1984). In it the mean is homogeneous without shifting the conditional mean function by the fixed effect, so that both the overall-constant and the time-invariant covariates are allowed in the model.
It should be noted that the longitudinal data employed in the current study are rather short-balanced panel with only two periods of observations for a given person. The presence of an individual-specific heterogeneity term would invalidate the assumption of independent sampling for the random effect model. On the other hand, the FENB model accounts for the latent heterogeneity and its correlatedness with the exogenous variables. Thus, it follows that the estimation results from the FENB model are much more robust and efficient. This is confirmed by the Schwartz Information Criterion and the Akaike information criterion, both of which suggest a preference for the FENB. In addition, the Hausman test also rejected the random specification and confirmed the evaluation that the FENB corresponds to the data. Thus, we focused on FENB in the following DD and DDD estimations.
As shown in Table 3, the estimated parameters for the two formulations of the control group differ slightly, although they tend to give the same qualitative results. In particular, both models estimated an increase in the expected number of doctor visits in the postreform period. The other results are common in the literature: men have fewer doctor visits than women, and the expected number of doctor visits is U shaped in relation to age. The less healthy individuals, with an MI-ID status and a higher chronic condition count, have more doctor visits than their counterparts. In addition, the NTG was observed to have more visits, while the ACG treatment group was found to have fewer visits relative to their counterparts.
However, the effect of the reform, as indicated by the interaction terms of the policy reform and the treatment group showed a substantial difference between these two specifications. In particular, the parameter for the specification of the natural control is significantly negative, while it is positive in that of the ACG control, although without statistical significance.
The results of the DDD regression specified in Equation 2 are also shown in Table 3. With the ACGPE being included among the regressors, the parameter for the interaction term of the policy reform and treatment group turned out to be positive, although insignificant, while the other coefficients showed the same effect for visits.
Results of DD and DDD Regressions.
Note. t statistics shown in parentheses. ACGPE = adjusted clinical group price elasticity; AIC = Akaike information criterion; BIC = Bayesian information criterion; DD = differences-in-differences; DDD = differences-in-differences-in-differences; EDCs = expanded diagnosis clusters; FENB = fixed-effects negative binomial; MI-ID = major illness identity.
p < .05. **p < .01. ***p < .001.
To control for individual behavioral idiosyncrasies, Table 4 shows the results of the average treatment effect using Equation 3 from the DDD regression. Each cell contains the mean difference between the average number of visits before and after the reform. In other words, it shows the SD for each group, along with the standard errors in parenthesis. It should be noted that all SD values of the pre-post difference are positive, except for that of the natural control when ACGPE <1, although it is statistically insignificant.
For individuals with ACGPE <1, it is unlikely that the prescription drug copayment plays a significant role because their price elasticity is rather insensitive. Thus, the SD when ACGPE <1 can be attributed to the variation in genuine health needs. This indicates an improvement (SD4 = −0.0372) and a deterioration (SD3 = 0.0692) in the natural control and natural treatment, respectively.
On the other hand, when ACGPE >1, as shown in Table 4, it is evident that the SD value of the natural control (SD2 = 0.214) is large and significantly positive, almost 2.7 times more than that of its counterpart (SD1 = 0.0799). As a result, it generates a decrease in doctor visits (DDACGPE>1 = SD1 − SD2), with a statistical significance when ACGPE >1. However, this decline is due mainly to the increase in the number of visits by the NCG, while the morbidity burden remains unchanged.
When controlling for the ACGPE of patients, we find that the average treatment effect in the treatment group (estimates in the left column of Table 4) is similar to that in the control group (estimates in the right column of Table 4). The number of doctor visits for individuals with an ACGPE >1 are higher than for their corresponding counterpart, that is, SD1 > SD3 and SD2 > SD4. However, the differences between ACGPE >1 and ACGPE <1 are minimal within the NTG but are substantial within the NCG. It is evident that the DD between individuals with ACGPE >1 and those with ACGPE <1 in the NCG is 25 times higher than that in the NTG (DDNCG = 0.2512 vs. DDNTG = 0.0107). This indicates that the potential moral hazard problem inherent in the exempt group is much more severe and distinct than that in the nonexempt group because the SD, when ACGPE <1 (SD3 and SD4), represents the change in visits based on genuine health needs.
Subtracting these two DDs using Equation 3 yields the average treatment effect of reform, as shown in Table 4; that is, DDD = DDtreatment − DDcontrol = DDACGPE<1 − DDACGPE<1. When compared with the previous DD results which did not control for the specific state of the ACGPE, the reform effect (DDD) is now also negative (−0.2410) but insignificant.
Predicted Average Treatment Effects (DD and DDD).
Note. Standard deviation shown in parenthesis. ACGPE = adjusted clinical group price elasticity; DD = differences-in-differences; DDD = differences-in-differences-in-differences.
*p < .05. **p < .01. ***p < .001.
Discussion
Robustness Checks
Because the DD approach relies on the important assumption of parallel trend, it is crucial to assess whether the treatment-assignment employed in this study, especially the constructed index of ACGPE, satisfies this requirement. We therefore performed a placebo test (Dumont, Fortin, Jacquemet, & Shearer, 2008; Galiani, Gertler, & Schargrodsky, 2005; Puhani & Sonderhof, 2010) for both the natural and ACG control.
As described in the previous section, the construction of an ACG control requires 2 years of data for computing price elasticity. Thus, it follows that 3 years of data are required for performing a placebo test—data from 2 years before the reform year t, that is, t − 2 and t − 1, are for the fake pre- and post-reform years, whereas the 2 years before the fake reform year, that is, t − 2 and t − 3, are necessary for calculating the ACGPE. However, data availability makes this difficult because the NHI was implemented in March 1996, and the complete data were collected from 1997 by NHRI, which leaves only 2 years of data before the reform. Thus, rather than making a comparison of 2 years before the reform using the same sample, we pooled Samples A and B, as described in the previous section, for conducting the placebo test to verify whether the ACG control satisfies the parallel trend assumption.
The placebo test was conducted in such a way that one sample was retained for the base year and the other one was employed for the fake reform year. For example, we tested a “placebo treatment effect” by pooling the data from Sample A in 1997 (pre-reform) with the data from Sample B in the same year (also pre-reform), but as if it was after the reform. The same placebo test could be applied to 1998. The coefficient of the placebo variable (=1 if treated in 1997 or 1998, respectively) should not be significant because no true reform was introduced during that year.
As shown in Table 5, the results indicate that the placebo coefficient is not significant, which not only confirms the validity of the natural control but also the proposed ACGPE for examining the impact of the reform.
Placebo Regressions.
Note. t statistics shown in parentheses. ACG = adjusted clinical group; ACGPE = ACG price elasticity; AIC = Akaike information criterion; BIC = Bayesian information criterion; EDCs = expanded diagnosis clusters; MI-ID = major illness identity; NTG = natural treatment group.
*p < .05. **p < .01. ***p < .001.
The advantages of employing the ACGPE as a comparison group are twofold. First, it separates the heterogeneous behaviors of the patients seeking medical care from their genuine need for health care. The selection bias is explicitly being taken care of in that the differential trend between individuals who are more susceptible to moral hazard versus those who are not is controlled. The individuals most likely to be affected by the increased copayment are those with ACGPE >1, while those with ACGPE <1 have no incentive to abuse the system. The ACGPE constructed in this study can complement the existing NCG by enhancing the comparability between treatment and control group using the DDD approach. Second, the constructed ACG control satisfies the parallel trend assumption and can thus serve as an ideal candidate for treatment assignment when there is a lack of an NCG.
Differences Between the Natural and Constructed Comparison Groups
The differences between the natural and the constructed comparisons groups before the reform can be illustrated with a 2 × 2 table. As shown in Table 6, it is clear that there is a 78.3% overlap in the two assignments, 77.3% of which is for both treatments and the remaining 1% is for both controls. The proportion of individuals with ACGPE >1 was larger in the NCG (93.2%) than in the NTG (89.9%). Individuals with an ACGPE <1 constituted 6.8% in the NCG and 10.1% in the NTG. At the same time, the sample ratio of the NTG and the control group was 90.2%–9.8% in the constructed comparison of ACGPE <1, and 85.6%–14.4% in the constructed counterpart of ACGPE >1. Altogether, this suggests that the heterogeneity in patient behavior inherent in the exempt group may be much more distinct than that in the nonexempt group. Since the exemption is determined mainly on socioeconomic factors, such as income and occupation, it is not surprising that the demand for health services is found to differ between these two groups (Newhouse, 1993). This difference in patterns could also reflect the differential in structural health conditions or the preferences between these two groups. In particular, individuals with ACGPE >1 that are exempt tend to have poorer health because their chronic condition count (1.96) as well as their proportion of major illnesses ID (0.02) are greater, but they have less medical utilizations at an average of 5.93 visits per person per year. On the other hand, individuals in the nonexempt group with ACGPE >1 are found to be healthier with a chronic condition count of 0.91 and a 1% major illness ID, but they have more visits (on an average of 8.27 visits per year). The discrepancies in demand behavior between the exempt and nonexempt groups may be attributed to the fact that medical care is to some extent perceived as a luxury good, which is consistent with the literature (Zweifel & Manning, 2000). Indeed, in most countries where copayment has been introduced, it has been observed that higher income people actually visited their physician more often than before the copayment was imposed (Cockx & Brasseur, 2003; Rosen et al., 2010; Stoddart, Barer, & Evans, 1993). This argument that higher income groups behave differently from lower income groups justifies employing the ACGPE to further scrutinize the DD estimator using the DDD approach.
The Overlap in the Two Specifications of Natural Control and ACG Comparison Group.
Note. Standard deviation given within parentheses. ACG = adjusted clinical group; ACGPE = ACG price elasticity; EDCs = expanded diagnosis clusters; MI-ID = major illness identity; NTG = natural treatment group.
Applications and Limitations
Our proposed method decomposed the reform effects into the effects of price on the patient’s behavior and their genuine need for health care. We therefore relied on the notion that the price elasticity of health care demand does not vary with the health status (Cockx & Brasseur, 2003; Manning & Marquis, 2001). Thus, we devised an index accordingly, that is, the ACGPE based on the ACG system in order to show how the frequency of visits to the doctor is decided, in comparison with the average price for their genuine health needs based on their ACG assignment. The constructed ACGPE represents the variation in a patient’s choice of visits relative to a reference price change of his or her ACG. The critical assumption therefore lies in whether the AVPACG truly represents the individual’s genuine health needs, so that it can serve as a reference in the denominator of the constructed ACGPE for comparing the variation in the patient’s visiting behavior in the numerator.
The overall goal for the ACG assignment is to identify groups of individuals with similar needs for health care resources and who share similar clinical characteristics (Johns Hopkins University Bloomberg School of Public Health, 2010). Thus, the denominator for the constructed ACGPE, that is, the variation rate of average ambulatory expenses per visit associated with a specific ACG, represents the change in morbidity burden of an individual with that of ACG. This can then be considered as a reference or benchmark for comparison with the increased/decreased rate of an individual’s choice of actual visits in the numerator for determining the ACGPE.
As noted earlier, patients may differ in their behavior of seeking health care, even with the same genuine health needs, as represented by the same ACG. Consider the following two extreme cases. First, imagine that John is generally healthy and has only a few visits for routine medical exams. He gets classified into an ACG that predicts his relatively low ambulatory cost based on his history of medical utilization and diagnostic information. Individuals with this ACG will have the same average price of ACG per visit, because the predicted ACG represents clinically logical categories for persons expected to require similar levels of health care resources. Then, in the second year, John breaks his arm. He now not only has several additional very expensive visits but will also be assigned a different ACG because he has a different etiology (ACG-2300) for that injury. The corresponding resources and thus the average price per visit for that ACG for delivering appropriate health care are accordingly higher. Under such circumstance, John’s ACGPE will most probably remain <1 because both the numerator (number of doctor visits) and the denominator (average price of the ACG) for his ACGPE increased.
On the other hand, consider Jane who was given diagnosis codes for uncomplicated diabetes mellitus and is therefore classified into the ACG for patients with stable, chronic medical conditions (ACG-0900), which will not change unless the number of different types of morbidity increases or decreases. Jane’s average price of her ACG per visit will remain the same owing to the fact that her diagnoses profile, that is, the morbidity burden, as predicted by the ACG, did not change. However, for some unobserved personal preference she decides to take preventive care by means of frequent doctor visits. In that case, her ACGPE will be >1 because her number of doctor visits is increased in the numerator, while the average price of her ACG remains the same in the denominator. Jane will be identified as being in the treatment group because her doctor visits are greater relative to her genuine needs for health care, as predicted by the average price of her ACG.
Thus, the ACGPE is designed in such a way that it can serve as a good tool for identifying patients who may deviate from the appropriate level of resource consumptions predicted by their ACG.
It should be noted that an implicit assumption must be made to validate the aforementioned analysis. In other words, there is no induced demand by physicians or collusions between providers and patients, such as upcoding. Other research will be needed if these factors are to be taken into account when evaluating the consequences of supply-side incentives on policy reform. It is beyond the scope of the current analysis.
Conclusion
Longitudinal data may provide a way to deal with unobservables, such as those in the DD approach and panel data regression methods. However, it relies on the assumption that the selection into treatment is not correlated with the observed or unobserved factors (confounders), which are associated with the outcome. Wherever this assumption is untenable, the inference will be contaminated with selection bias due to a failure to control for important, but unobserved or unobservable characteristics. The assignment of treatment should control for differences in the outcomes between treatment and control groups, which, in the absence of any reform, are assumed to be constant over time. Therefore, if there is a difference in the outcomes caused by the individual’s heterogeneous behavior, such as moral hazard, then this difference should be identified prior to the reform and be incorporated as a genuine counterfactual to assess the effects of the policy. If not, then the estimated reform effect becomes an illusion and thus misleading. As the NCG is most often a very selective group with low socioeconomic standing, the effect of the reform may be over- or underestimated, depending on how much or how little it is influenced by the confounders. Combining the ACGPE with the DDD approach allows us to control for unobserved time-varying components and increase the credibility of the identification of the treatment-effect accordingly.
Footnotes
Acknowledgments
This article is based in part on data from the National Health Insurance Research Database provided by the Bureau of National Health Insurance, Department of Health, Taiwan, and managed by the National Health Research Institute. The interpretation and conclusions presented herein do not represent those of the Bureau of National Health Insurance, Department of Health, nor the National Health Research Institute in Taiwan.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The National Science Council, Taiwan, under grant number NSC99-2410-H-327-002-SS2.