Effects of Recent Carrot-and-Stick Policy Initiatives on Private Health Insurance Coverage in Australia

Transcription

1 THE ECONOMIC RECORD, VOL. 81, NO. 254, SEPTEMBER, 2005, Effects of Recent Carrot-and-Stick Policy Initiatives on Private Health Insurance Coverage in Australia ALFONS PALANGKARAYA and JONGSAY YONG Melbourne Institute of Applied Economic and Social Research, Centre for Microeconometrics, University of Melbourne, Melbourne, Australia The Australian government implemented a sequence of new policies during and raised the take-up rate of private health insurance (PHI) by 35 per cent. Because they were implemented sequentially, their individual effectiveness is not quite clear. We isolate the effects of Lifetime Health Cover (LHC) introduced at the last stage of the sequence using a counterfactual analysis of PHI demand with and without the new policies. Unlike earlier studies which attributed the bulk of the increase in PHI coverage to LHC, we find LHC may only account for as low as 42 per cent and no more than 75 per cent of the increase. I Introduction The Australian health-care system is based on a universal access principle, under which every person regardless of income and health risk is covered by Medicare, a publicly funded, universal health insurance scheme. However, private health insurance (PHI) has always been a prominent feature of the Australian health system, despite the availability of Medicare. 1 For much of the 1990s, some per cent of the The authors acknowledge helpful comments from participants of the Melbourne Institute seminar and the 2004 Australian Meeting of the Econometric Society on an earlier version of this paper. We also thank Harry Bloch and two anonymous referees. Remaining errors are the authors sole responsibility. Jongsay Yong acknowledges the financial support from the University of Melbourne Faculty Research Grant. JEL classifications: I10, I18 Correspondence: Alfons Palangkaraya, Melbourne Institute of Applied Economic and Social Research, Centre for Microeconometrics, University of Melbourne, 161 Barry Street, Parkville, Vic. 3010, Australia. alfonsp@unimelb.edu.au 1 The first Australian universal health insurance system, called Medibank, was announced in 1972 and formally in place in Following some major revamps, the name was later changed to Medicare in population was covered by PHI. Until recent years, however, the coverage of PHI was on a declining trend since the implementation of Medicare (or Medibank before 1984) in There are a variety of reasons, not least of which is the appeal of Medicare/Medibank to the masses. With a purported goal of reducing the burden on public hospitals, the Australian government implemented a sequence of new policies during The three major policy initiatives were, in chronological order: (i) the Private Health Insurance Incentives Scheme (PHIIS), introduced in 1997, which imposes a tax levy on high-income earners who do not have PHI, and a means-tested subsidy scheme for low-income earners who purchase PHI; (ii) a 30 per cent premium rebate, introduced in 1999, to replace the means-tested subsidy under PHIIS the 30 per cent rebate is nonmeans tested, and applies to all PHI policies, including existing ones that were already in place; and (iii) Lifetime Health Cover (LHC), introduced in 2000, permits a limited form of age-related risk rating by PHI funds. Under LHC, insurance funds are allowed to discriminate consumers by age at the time of entry. The immediate aim of these policies was to raise the take-up rate of PHI. As a result of these policies, the proportion of the population with private hospital cover increased from 31 per cent in 1999 to more C The Economic Society of Australia. ISSN

2 2005 EFFECTS OF HEALTH INSURANCE POLICY 263 than 45 per cent in 2001, an increase of more than 14 percentage points in 2 years. Two components of the new policies, the 30 per cent premium rebate and LHC, have been widely regarded as the most effective in raising the PHI take-up rate. It is not clear, however, what proportion of the increase can be attributed to each of these policies. Butler (2002) examines the trend in the proportion of population with private hospital insurance cover between June 1984 and March By noting the timing of the sequential policies introduced, beginning in mid-1997, he argues that it was LHC that induced the bulk of the increase in PHI take-up rate. Using a similar policy timing idea but with a more rigorous trend analysis, Frech et al. (2003) attempt to measure the relative impact of the different policies. They estimate that the 30 per cent rebate led to an 11 per cent increase in PHI demand. However, they are not able to disentangle the effects of the various policies. The question is of interest because of the substantial cost differences between LHC and the 30 per cent rebate. The cost effectiveness of each policy has attracted heated debate among academics and policy-makers. 2 This paper departs from previous studies by using microlevel cross-section data to disentangle the effects of the two policies. We argue that although a significant proportion of the increase in PHI coverage occurred after the introduction of LHC, this policy was introduced on top of, not in place of, the 30 per cent premium rebate and the tax levy under PHIIS. This means that the separate contributions of these policies cannot be readily inferred from the time-series data. Using cross-sectional microlevel data, this paper contributes to the discussion by providing separate estimates of the effects of LHC and the other PHI policies. A significant improvement over previous studies comes from our use of a microlevel model of PHI demand, which allows us to take advantage of the design of the two policies to separate their effects. Using unit record data from the 1995 and 2001 National Health Survey (NHS) data, we estimate PHI demand equations before and after the implementation of the new policies, separately for single individuals and families. This allows us to construct a hypothetical or counterfactual scenario, as in Barrett and Conlon (2003), to measure the effects of the recent PHI policies on PHI coverage by accounting for those who are unlikely to have PHI without these policies. Furthermore, by using the age criterion of the LHC policy, we are able to isolate the contribution of LHC. For singles, 2 It is estimated that the 30 per cent rebate costs more than $A2 billion per year, while LHC costs practically nothing to the government (see e.g. Duckett & Jackson, 2000). we find that LHC accounts for at least 42 per cent and at most 61 per cent of the increase in PHI coverage. For families, the corresponding figures are 42 and 78 per cent of the increase, respectively. The effects of LHC, expressed as a proportion of the total population, are between 42 and 75 per cent of the estimated increase in PHI coverage due to the combined effects of all recent PHI policies. Although these figures show the significant impact of LHC on PHI coverage, they are much lower than those suggested by previous studies. It is perhaps instructive to contrast our approach with that of the earlier time-series based studies. In the latter, a time trend is usually estimated and the policy effects are usually denoted using a time dummy variable, which shifts the intercept or the slope of the estimated equation. Extending this approach to allow for interactions of different policies requires one to interact the policies in some specific form, for example, multiplicatively and the various policy effects are identified by deviations from the time trend. In contrast, our approach separates the policy effects by exploiting the policy design of LHC, under which only individuals and families in the target age groups are affected. The crucial identifying assumption is that LHC does not affect households in the non-target groups, while other non-lhc, policies affect all households in a uniform manner. The rest of this paper proceeds as follows. In Section II, we provide a brief discussion on the changes in the PHI policies. This is followed by the empirical model specifications and an explanation of the data and variable construction in Section III. In Section IV we estimate the policy effects, from which we decompose the effects of LHC. Section V concludes the paper. II PHI Policy Changes There were three major sets of PHI policy changes implemented between 1997 and 2000 that may affect PHI demand. The first, known as PHIIS, was introduced in 1997 and included a means-tested PHI premium subsidy and a 1 per cent Medicare tax surcharge for high-income earners who did not have an approved PHI policy. Then, in 1999 the means-tested rebate was replaced with a non-means tested 30 per cent premium rebate for all PHI policies, including insurance policies that were already in place before the rebate was announced. Finally, introduced in 2000 was LHC, which allows health funds to discriminate across consumers according to their age of entry into the fund (see details below). Taken together, these policy changes have been effective in raising the coverage of PHI. As shown in Figure 1, the percentage of

3 264 ECONOMIC RECORD SEPTEMBER Per cent pop. insured FIGURE 1 Proportion of Population with Private Health Insurance, % 45% 40% 35% 30% PHIIS 30% Rebate LHC grace period LHC in force (15 July 2000 ) No. families (individuals) FIGURE 2 Reasons for having Private Health Insurance family single 25% Jun- 96 Dec- 96 Jun- 97 Dec- 97 Jun- 98 Dec- 98 Jun- 99 Dec- Jun Dec- 00 Jun- 01 Dec- Jun Dec- Jun population covered by PHI rose from 31 per cent in 1999 to a high of 45 per cent at the end of From the timing of the policies, and the continuing decline in PHI coverage after the introduction of PHIIS, it is widely accepted that the 30 per cent premium rebate, and particularly LHC, are the most effective policies. Therefore, these are the focus of a majority of earlier studies. The purpose of the LHC scheme is to induce the low-risk population back into the PHI system. It is a response to the finding that lowrisk individuals were leaving the PHI system, resulting in a system consisting mostly of high-risk individuals a phenomenon referred to as the adverse selection spiral (see e.g. Industry Commission, 1997). Consequently, the scheme is designed around a financial penalty for the low-risk groups. The target group is individuals between 30 and 65 years of age. The amount of the penalty is set at 2 per cent of the base premium for each year beyond the age of 30 years for anyone in the targeted population taking up PHI for the first time. The policy was announced in September 1999, but took effect in July 2000 to allow for a 9-month grace period for people with no PHI to purchase one and avoid paying the higher premiums. Figure 1 shows an appreciable rise in PHI coverage immediately following the introduction of LHC. Hence, one is tempted to attribute most of the increase to LHC. However, there are several plausible reasons why the effects of the 30 per cent rebate should not be discounted entirely. First, from the 2001 NHS, approximately 12 per cent of the respondents aged between 30 and 65 years regarded LHC as one of the reasons for having PHI (see Fig. 2). Second, the same data reveal that this age group accounts for approximately 72 per cent of the PHI population. This means that No. families (individuals) 0 Security Private Choice Wait Always had Ancillary Medicare levy LHC Elderly Other Reasons Other f inancial Condition FIGURE 3 Reasons for not having Private Health Insurance Can't afford High risk Not worthy Medicare Don't need No Medicare levy Gap fees Self-insured Concession Low priority Other Reasons family single at least 28 per cent of the insured population was not affected by LHC. Finally, the consumer theory asserts that what matters in the end is whether having PHI is optimal, given an individual s budget constraint. Indeed, Figure 3 shows that affordability is a major consideration in the consumers PHI decision. It is quite plausible that the rise in PHI coverage following the introduction of LHC would have been much smaller if PHI was still unaffordable relative to other goods in the absence of the 30 per cent rebate.

4 2005 EFFECTS OF HEALTH INSURANCE POLICY 265 III Empirical Framework and Data The specification of our empirical demand model follows Cameron and Trivedi (1991) and is similar to other Australian studies (e.g. Hopkins & Kidd, 1996; Barrett & Conlon, 2003). The basic framework is motivated by the theory of consumer choice under uncertainty, where households only purchase PHI if the expected utility from purchasing exceeds that from not purchasing. The decision to purchase PHI can thus be specified as a discrete choice model. That is, defining the binary variable PHI as an observed random variable with value 1 if PHI is chosen and value 0, otherwise. The probability of purchasing PHI is: Prob[PHI = 1] = f (x,β) + e, (1) where f is a cumulative distribution function, x is a vector of observable individual characteristics, β is the corresponding vector of parameters to be estimated and e is a random term representing the unobservables. We specify f so that equation 1 can be estimated as a standard logit model. Following the three studies cited earlier, x contains variables that capture variations in: (i) expected medical needs (gender, family size, age and doctor visits); and (ii) risk aversion, preference as well as other socioeconomic background (smoking, income, types of employment, residential area and education). Missing from the list is the rate of insurance premium, which unfortunately is not available from existing Australian datasets. We estimate equation 1 twice, once using the 1995 NHS data and once using the 2001 NHS data. 3 The interview of the 2001 survey took place between February and November Therefore, it ought to capture most, if not all, of the effects of the PHI policies mentioned earlier. In their original formats, the two datasets contain and unit records for 2001 and 1995, respectively. Since the PHI decisions of singles and families are plausibly driven by different considerations, we split the data accordingly and estimate equation 1 separately for singles and families for each survey. In addition, since the 2001 survey only randomly interviewed one adult from a family type household, while the 1995 survey interviewed all adults in the sampled households, we randomly select one adult from households in the 1995 data. This is to ensure that the two samples are as comparable as possible. However, we note that the age variable of 3 The 2001 NHS data are available from the enhanced Confidentialised Unit Record Files, the data can only be accessed by the Australian Bureau of Statistics (ABS) remote access protocol known as the Remote Access Data Laboratory (RADL) through the ABS website. families in our sample serves only as a proxy, it does not allow precise identification of the age criterion under LHC, which consists of the ages of both parents of a family. Finally, there are several data-cleaning steps that we undertook. These include dropping full-time students, non-family members, age <18 years for singles and age <20 years for families, and singles age <22 years who were covered by family PHI. After undertaking these steps, we end up with useable records of, respectively for the 1995 and 2001 NHS, 4648 and 4710 singles and 7059 and 9543 families. Table 1 lists the variables used in the estimation of equation 1. Note that all person description variables in the case of family units refer to the characteristics of the selected adult of the particular family. In addition to the variables listed in Table 1, we also include 13 regional dummy variables, which denote whether the individual or family s residence is in a metro or rural area in each state/territory, and 13 age dummy variables, with value 1 if the age of the individual or the family s selected adult falls in the indicated age interval. Table 2 summarises the estimated coefficients of the PHI demand model as specified in equation 1, for singles and families, and for the 1995 and 2001 NHS data. As can be seen from the table, most of the usual determinants of PHI demand are significant and have the expected signs. Demand is increasing in income, but this effect is diminishing as suggested by the negative sign of the squared income term. This probably captures the preference for private health care and at the same time the ability to afford the costs. Comparing the coefficients of 1995 and 2001, we see that the signs of estimated coefficients are mostly as expected and consistent in both years. The coefficients for gender, income, education, occupation, risk aversion and health risks are quite stable in both years. It is worth noting that there is no variable which shows a reversal in sign and yet remains statistically significant in both years. Contrary to what one would expect from adverse selection, smokers, whether single or in a family, are associated with significantly lower probabilities of having PHI in both years. One possible explanation is perhaps this reflects the lower risk aversion of smokers, and it dominates their higher health care needs consideration. The signs of the estimated coefficients for doctor visits and the number of chronic conditions are consistent with usual expectations in that they capture the expected medical needs in the future. One particularly interesting finding from the 1995 family estimates is that the coefficients on doctor visits seem to suggest

5 266 ECONOMIC RECORD SEPTEMBER TABLE 1 List of Variables in Logit Estimation Dependent variable PHI Explanatory variables female immig govcrd stdinc stdinc2 famsize childlt6 degree diploma postscd admin trade clrksrv prof paraprof plantop drvisit1 drvisit2 drvisit3 drvisit4 chdrvis1 4 chnum chnumcld smoker Private health insurance (1 = yes, 0 = no) 1 = if hospital cover (with or without ancillary) is purchased (for family, 1 = if at least one family member has hospital cover) Dummy, 1 = female (not applicable for family units) Dummy, 1 = immigrant Dummy, 1 = government concession/entitlement card holder Standardised value of annualised weekly income ( = actual mean income) squared standardised income Number of persons living in the household (not applicable for singles) Number of children of age 6 years (not applicable for singles) Dummy, 1 = has a Bachelor degree or higher Dummy, 1 = has an associate diploma Dummy, 1 = has basic, skilled and/or other post-secondary qualifications Dummy, 1 = occupation is in an administrative position Dummy, 1 = occupation is trade Dummy, 1 = occupation is clerical or services Dummy, 1 = occupation is in the professional category Dummy, 1 = occupation is paraprofessional Dummy, 1 = occupation is categorised as plant operator Dummy, 1 = has a doctor visit within last 2 weeks Dummy, 1 = has a doctor visit within last 2 weeks to 3 months Dummy, 1 = has a doctor visit within last 3 6 months Dummy, 1 = has a doctor visit within last 6 12 months Defined similarly as drvisit1 4 above for children (not applicable for singles) Number of long-term chronic conditions Average number of long-term chronic conditions of all children (not applicable for singles) Dummy, 1 = current regular smoker that the expected medical needs of children are more important than those of adults. By comparing the 1995 and 2001 coefficients, we can obtain an indication of the risk pool of the insured. If the policies induce more individuals from the lower risk group to purchase PHI, the importance of the riskrelated variables should be lower in The doctor visits variables of singles and families, both adult and children, seem to indicate this is the case. Except for drvisit3 and drvisit4 of families, the magnitudes of the estimated coefficients decrease between the two surveys. However, the other risk-related variables (e.g. number of chronic conditions and smoking) seem to provide contradicting evidence. It should, however, be noted that the doctor visit variables may be subject to endogeneity bias, in that households with PHI are more likely to have a higher number of doctor visits. However, since our main concern is to estimate the difference in predicted probabilities as outlined below, so long as the same bias exists for both 1995 and 2001 estimates, our results are unlikely to be severely affected. We further note that since all other policies except LHC are age-independent, on average, the distance between the LHC target group (age years) and the LHC non-target group (age and 65+ years) would increase. In other words, since the base age group (age years) belongs to the LHC target group, the coefficients of the non-target group would be lower in The evidence, although limited, seems to support this point. IV Estimated Policy Effects Based on the estimated coefficients, we estimate the increases in PHI membership that result from the combination of policy initiatives introduced between 1995 and We do this by creating a hypothetical or counterfactual situation by applying the 2001 NHS data on the 1995 coefficient estimates. The resulting probabilities are called the hypothetical predicted probabilities, denoted Prob 2001/1995 i, for household i. In arriving at the hypothetical predicted probabilities, we in effect force households in 2001 to apply the

6 2005 EFFECTS OF HEALTH INSURANCE POLICY 267 TABLE 2 Coefficient Estimates of Logit Private Health Insurance Demand Models Single Family Explanatory variables female (0.0854) (0.0803) famsize (0.0430) (0.0356) childlt (0.0911) (0.0750) nochild (0.1012) (0.0918) immig (0.0992) (0.8624) (0.0646) (0.3985) govcrd (0.1573) (0.1338) (0.0826) (0.0780) stdinc (0.0867) (0.0814) (0.0389) (0.0495) stdinc (0.0285) (0.0217) (0.0233) (0.0256) vicmet (0.1331) (0.1311) (0.1008) (0.0928) qldmet (0.1960) (0.1649) (0.1413) (0.1072) samet (0.1449) (0.1431) (0.1115) (0.1113) wamet (0.1733) (0.1443) (0.1274) (0.1048) tasmet (0.3384) (0.2115) (0.2341) (0.1663) nt (0.1789) (0.2967) (0.1331) (0.2159) act (0.1663) (0.1619) (0.1288) (0.1115) nswrur (0.2558) (0.1612) (0.1634) (0.1072) vicrur (0.2143) (0.1852) (0.1481) (0.1182) qldrur (0.1787) (0.1561) (0.1298) (0.1028) sarur (0.2587) (0.2298) (0.1571) (0.1541) warur (0.3529) (0.2548) (0.1875) (0.1528) tasrur (0.2279) (0.2152) (0.1720) (0.1385) age (0.2556) (0.3916) age (0.1870) (0.2070) (0.1636) (0.1747) age (0.1949) (0.1883) (0.1274) (0.1155) age (0.2089) (0.1846) (0.1126) (0.1024) age (0.2156) (0.1927) (0.1051) (0.0927) age (0.2220) (0.1882) (0.1094) (0.1006) age (0.2388) (0.1839) (0.1278) (0.1162) age (0.2526) (0.1963) (0.1450) (0.1321) age (0.2549) (0.2020) (0.1601) (0.1385) age (0.2560) (0.2110) (0.1691) (0.1501) age (0.2619) (0.2065) (0.1764) (0.1538) age (0.2589) (0.2120) (0.2138) (0.1691) age80p (0.2594) (0.2055) (0.2747) (0.2070) degree (0.1372) (0.1252) (0.1021) (0.0884) diploma (0.1479) (0.1344) (0.0999) (0.0873) postscd (0.1003) (0.0905) (0.0700) (0.0603) admin (0.2132) (0.2176) (0.1181) (0.1248) trade (0.1702) (0.1907) (0.1112) (0.1045) prof (0.1867) (0.1727) (0.1197) (0.1009) paraprof (0.2259) (0.1827) (0.1386) (0.1036) clrksrv (0.1425) (0.1412) (0.0846) (0.0770) plantop (0.2478) (0.2063) (0.1451) (0.1183) drvisit (0.1390) (0.1305) (0.0968) (0.0845) drvisit (0.1304) (0.1261) (0.0896) (0.0812) drvisit (0.1448) (0.1418) (0.1013) (0.0900) drvisit (0.1537) (0.1522) (0.1070) (0.0944) chnum (0.0196) (0.0259) (0.0149) (0.0172) chnumcld (0.0460) (0.0422) chdrvis (0.1754) (0.1555) chdrvis (0.1637) (0.1417)

7 268 ECONOMIC RECORD SEPTEMBER Single TABLE 2 Continued Family Explanatory variables chdrvis (0.1867) (0.1550) chdrvis (0.1994) (0.1740) smoker (0.0940) (0.0885) (0.0698) (0.0614) constant (0.2494) (0.8895) (0.2178) (0.4356) LLF Pseudo R N PHI = 1 25% 37% 44% 52% Note: Figures in parentheses are standard errors. decision rules as embodied in the estimated 1995 coefficients. The rationale being that the estimated 1995 coefficients should not reflect the effects of recent PHI policies. Thus, for each household in the sample, we are able to deduce its PHI decision under the hypothetical scenario, which is constructed to abstract from any policy effect. We compare the household s hypothetical PHI decision to its actual decision in A household is said to be affected by the PHI policies if it had no PHI under the hypothetical scenario but was actually covered by PHI in The estimated policy effects (EPE) are simply the proportion of households in the sample who were affected by the PHI policies. The specifics are as follow. First, we define an indicator variable, HPHI i, for household i in the sample to denote the household s insurance decision under the hypothetical scenario: HPHI i = { 1 if Prob 2001/1995 i c 0 otherwise, (2) where c serves as a cut-off point, which is set at the actual 2001 sample proportion, as suggested by Cramer (1999). That is, we regard a household as likely to have PHI under the hypothetical scenario if its hypothetical predicted probability is higher than the sample proportion of households who had PHI in Next, we define another indicator variable, I i, to denote whether household i s PHI decision is affected by the PHI policies. { 1 if PHIi = 1and HPHI i = 0 I i = (3) 0 otherwise, where PHI i is household i s actual PHI status in Let N j and n j be, respectively, the set and number of households in age group j. Then, the EPE on age group j is simply: EPE j = I i Nj i, (4) j n j where j n j is the total number of households across all age groups. Recall that we distinguish two types of households in our sample: singles and families. Thus the above exercise is performed for each household type separately. The results are summarised in Table 3. The last column of the table shows that 36.9 per cent of single individuals had PHI in 2001, and this consists of 21.4 per cent who would have had PHI in 2001 even if there were no recent policy initiatives, and the other 15.5 per cent took up PHI in response to the recent policies. Similarly for families, 51.9 per cent had PHI in 2001, and this can be decomposed into 21.8 per cent who would have had PHI even if there were no recent policies and the remaining 30.1 per cent who responded to the recent policies. Columns 2 4 of Table 3 provide the breakdown of the above figures into three age groups: group 1 (18 29 years), group 2 (30 69 years) and group 3 (70 years and more). Groups 1 and 3 are, by policy design, not affected by LHC; thus, for singles, at least 6.1 ( ) out of the 15.5 percentage points are not related to LHC. The corresponding figure for families is 6.6 out of 30.1 percentage points. Column 3 shows that, out of the 15.5 percentage points of singles who took up PHI in response to the new policies, approximately 9.4 percentage points (or 61 per cent of 15.5) were from group 2, the LHC target group. We also compute 95 per cent confidence intervals for the

8 2005 EFFECTS OF HEALTH INSURANCE POLICY 269 TABLE 3 Estimated Proportion of Population Affected by Private Health Insurance (PHI) Policies Group 1: age years Group 2: age years Group 3: age 70+ years All Singles Total If no policy change Estimated policy effects % confidence interval (4.2, 15.2) (6.8, 24.6) Families Total If no policy change Estimated policy effects % confidence interval (15.8, 30.2) (20.8, 37.2) All figures are in percentage points. For families: years. TABLE 4 Marginal Effects of Age on Probability of having Private Health Insurance (PHI) in 2001 Single Family Marginal effects Prob[PHI = 1] Marginal effects Prob[PHI = 1] Reference group Age years Group 1 Age years Age years Age years Average predicted probabilities Group 2 Age years Age years Age years Age years Age years Age years Age years Average predicted probabilities Group 3 Age years Age years Age 80 years and more Inclusive of reference age group. EPE on group 2 and the total sample using the delta method. 4 We next attempt to attribute the EPE on group 2 to LHC and other non-lhc, PHI policies, chiefly PHIIS and the 30 per cent premium rebate. We do this by 4 The delta method is used to compute the standard errors, so as to construct the confidence intervals (see Appendix I for details). taking into account the marginal effects of the age variables. Table 4 presents the average marginal effects of all age variables with respect to the base age group (age years) in In addition, the table also gives the predicted probabilities for the base age group so that the predicted probabilities of all other age groups can be inferred directly from their marginal effects. These age-dependent variations in predicted probabilities allow the effects of LHC to be disentangled from other effects.

9 270 ECONOMIC RECORD SEPTEMBER TABLE 5 Estimated Lifetime Health Cover (LHC) Contributions Singles Families Group 2 APBs Estimated non-lhc effects Estimated LHC effects As per cent of group 2 APBs 69% 53% APB, average predicted probabilities. Instead of giving a point-estimate, we attempt to provide a lower and an upper bound of the LHC effects. We first derive the lower bound. Although both groups 1 and 3 are not affected by LHC, we ignore group 3 in our analysis, as it mostly consists of retired individuals with high health risk and thus are non-comparable with the other two groups. Controlling for all other observable variables, we regard the average predicted probabilities (APBs) of group 1 as reflecting the policy effects other than LHC. Therefore, we will use the difference of the APBs between groups 1 and 2 to arrive at the lower bound estimates. From the within group averages in Table 4, the APBs of having PHI for singles and families in group 2 are and , respectively. As discussed earlier, these predicted probabilities consist of the effects of LHC and other policy effects. To disentangle these effects, we make use of the APBs of group 1. Since by design group 1 is not affected by LHC, their APBs provide an estimate of the effects of other non-lhc, PHI policies. Assuming different age groups respond in the same way to non-lhc policies, we can treat the difference between the APBs of groups 1 and 2 as the policy effects attributable to LHC. Thus, LHC effects = APBs group 2 APBs group 1. (5) The results are summarised in Table 5. The estimated LHC effects are and for singles and families, respectively. Expressed as a percentage of average probabilities, LHC accounts for approximately 69 and 53 per cent of the predicted probabilities of having PHI for singles and families in group 2, respectively. Applying those percentage estimates to the estimated overall policy effects shown in Table 3, we provide an estimate of the lower bound of the LHC effects in terms of the proportion of population with PHI. The results are presented in Table 6, which we construct in the following manner. We first reproduce the estimated overall policy effects in the first row of Table 6. Using the percentage estimates of 69 per cent in Table 5, we arrive at 6.4 (69 per cent of 9.4) percentage points as the estimated effects of LHC on singles in group 2. The same reasoning yields 12.5 percentage points as the estimated effects of LHC on families in group 2. We treat these figures as lower bounds of the LHC effects because the analysis is done by assuming that LHC does not affect the other two age groups. The upper bounds in Table 6 are computed by simply attributing all EPE on group 2, the LHC target group, to LHC. In other words, we regard other policies as completely ineffective in increasing the PHI take-up rate of group 2. In contrast, for groups 1 and 3 which are not affected by LHC, we attribute all the EPE to the other, non-lhc, policies. We note that the true effects are probably much lower than these upper TABLE 6 Estimated Private Health Insurance (PHI) Policy Effects in 2001 Group 1: age years Group 2: age years group 3: age 70+ years All Singles Overall policy effects Attributed to other policies Attributed to LHC Lower bound Upper bound Families Overall policy effects Attributed to other policies Attributed to LHC Lower bound Upper bound All figures are percentage points. For families: years.

10 2005 EFFECTS OF HEALTH INSURANCE POLICY 271 bounds since they are obtained on the assumption that individuals and families in group 2 responded only to LHC and not other policy initiatives. 5 We make three remarks. First, we conjecture that the true magnitude of the total effects would be closer to the lower than the upper bounds. This is because the potential spillover effects of LHC on groups 1 and 3, say because of confusion in understanding the regulations, are most likely to be small. In contrast, the potential bias due to other non-lhc related, agecorrelated factors that we ignored in estimating the upper bounds can be large. Second, it is worth noting that a weighted average scheme can be used to obtain an interval estimate for the total effect of LHC on the whole population. Since approximately 20 per cent of the respondents in the 2001 NHS are singles, thus we use a 1:4 ratio. Applying this population weight, we estimate that LHC accounts for between 42 and 75 per cent of the rise in PHI membership in the Australian population. For the same reasons mentioned earlier, we think the true value is likely to be much closer to the former than the latter. Third, it is possible to construct confidence limits for the upper- and lowerbound estimates by making use of the confidence limits in Table 3 and following exactly the same procedure as outlined earlier. V Conclusion Economic efficiency alone dictates that if a policy costs less and yields higher desired effects, then it should be a preferred option. Among the recent Australian PHI policy initiatives, the LHC policy costs almost nothing to the government while the 30 per cent premium rebate costs more than $A2 billion per year. The crucial question is whether or not the latter, a much more expensive policy, is as ineffective as claimed by many authors and policy commentators. Mindful of these arguments, we proceed to disentangle the effects of the LHC from other policy initiatives. We do so by estimating PHI demand models using microlevel data before and after the policies were in place. The main findings are that LHC accounts for at least 42 per cent and at most 75 per cent of the overall rise in PHI coverage. We also argue that the true share of LHC is probably much closer to the lower bound. Thus, we conclude that the contribution of the 30 per cent premium rebate and the Medicare tax levy is far more substantial than widely believed. 5 An age-based difference-in-difference or a regression discontinuity analysis may be used to arrive at a more accurate estimate of the LHC effect. However, the difficulty is in defining the age group of families (see e.g. Palangkaraya & Yong, 2004). We note that our results are obtained by exploiting the policy design of LHC, in that only certain age groups are affected by the policy. In addition, to disentangle different policy effects, we further assume that other, non-lhc, PHI policies affect all age groups in a uniform way. There are several possibilities of how this study can be improved. First, in modelling PHI demand, we do not consider the fact that PHI in Australia is secondary to a freely available alternative, the Medicare. A more careful modelling of such duplicate cover, along the lines of Vera-Hernandez (1999), may result in more precise estimates. Another important improvement can be made by carefully taking into account the effects of the 1 per cent Medicare levy on high-income earners. This can be done, for example, by using the regression discontinuity approach by exploiting the fact that the levy only applies if a specific income threshold is reached. This would result in a more precise accounting of the contributions of each policy initiative. Finally, we note that our approach is different from the Oaxaca Blinder decomposition, which is a technique of decomposing intergroup differences into those due to different observable characteristics and those due to different effects of coefficients of groups. In our context, it is conceptually possible to use the Oaxaca Blinder decomposition along the line suggested by Fairlie (2003). However, this requires us to pool the 1995 and 2001 NHS data. Unfortunately, this is not possible due to the particular access protocol imposed on the enhanced version of the 2001 NHS data by the Australian Bureau of Statistics. REFERENCES Barrett, G.F. and Conlon, R. (2003), Adverse Selection and the Decline in Private Health Insurance Coverage in Australia: , The Economic Record, 79, Butler, J.R.G. (2002), Policy Change and Private Health Insurance: Did the Cheapest Policy Do the Trick, Australian Health Review, 25, Cameron, A.C. and Trivedi, P.K. (1991), The Role of Income and Health Risk in the Choice of Health Insurance: Evidence from Australia, Journal of Public Economics, 45, Cramer, J.S. (1999), Predictive Performance of Binary Logit Models in Unbalanced Samples, The Statistician, 48, Ducket, S.J. and Jackson, T.J. (2000), The New Health Insurance Rebate: An Inefficient Way of Assisting Public Hospitals, Medical Journal of Australia, 172, Fairlie, R.W. (2003), An Extension of the Blinder-Oaxaca Decomposition Technique to Logit and Probit Models,

11 272 ECONOMIC RECORD SEPTEMBER Discussion Paper No. 873, Economic Growth Center, Yale University. Frech III, H.E., Hopkins, S. and Macdonald, G. (2003), The Australian Private Health Insurance Boom: Was it Subsidies or Liberalised Regulation?, Economic Papers, 22, Greene, W.H. (2003), Econometric Analysis, 5th edn, Prentice-Hall, Englewood Cliffs, NJ. Hopkins, S. and Kidd, M.P. (1996), The Determinants of the Demand for Private Health Insurance under Medicare, Applied Economics, 28, Industry Commission (1997), Private Health Insurance, PC Report No. 57. Available from URL: privatehealth.pdf Palangkaraya, A. and Yong, J. (2004), How Effctive is Lifetime Health Cover in Raising Private Health Insurance Coverage in Australia? A Regression Discontinuity Approach, Working Paper No. 33/2004, Melbourne Institute of Applied Economic and Social Research, University of Melbourne. Vera-Hernandez, A.M. (1999), Duplicate Coverage and Demand for Health Care: The Case of Catalonia, Health Economics, 8, Appendix I: Constructing Confidence Limits We outline, in three steps, how we construct the 95 per cent confidence limits for the EPE in Table 3, on the proportion of population with PHI: 1. For each observation in the 2001 sample, we apply the delta method (Greene, 2003, section D.2.7) on the estimated function (equation 1) and the variance covariance matrix of the coefficientestimators to obtain the standard errors for the counterfactual predicted probabilities Prob 2001/1995 and construct the corresponding confidence interval. The averages of these counterfactual probabilities (by age) and their confidence limits along with the actual predicted probabilities are shown in Figures A1 and A2, for singles and families, respectively. 2. We convert the lower and upper limits obtained in the first step into lower and upper hypothetical indicators of PHI ownership in 1995 using the same procedure and probability cut-off point for deriving the point-estimates (HPHI i ) as described in Section IV (see equation 2). FIGURE A1 Average Predicted Probabilities of having PHI, Single 2001 Probability Coeff Coeff. 95% Con. Int Age (years) FIGURE A2 Average Predicted Probabilities of having PHI, Family 2001 Probability Coeff Coeff. 95% Con. Int Age (years) 3. We compare each of the lower and upper estimates of HPHI i to the actual 2001 PHI status to obtain the limits of the policy effects estimates shown in Table 3. The procedure is the same as described in Section IV (see equations 3 5).