Number of drugs provided by the Pharmaceutical Benefits Scheme and mortality and hospital utilization in Australia, 2002–2019

Abstract

We analyze the relationship between long-run changes in the drugs provided by the Pharmaceutical Benefits Scheme (PBS) and mortality and hospital utilization in Australia, by analyzing the correlation across diseases between the change in the number of drugs used to treat the disease provided and the subsequent change in mortality or hospital utilization from that disease.

Our estimates indicate that diseases for which there were larger increases in the number of PBS drugs tended to have smaller subsequent growth in premature (before ages 85, 75, and 65) mortality. Diseases for which there was larger growth in the number of PBS drugs also tended to have smaller growth in the number of hospital days 2–10 years later. The reduction in the number of hospital days appears to be primarily attributable to a reduction in average length of stay.

We estimate that the 1996–2013 increase in the number of PBS drugs was associated with a reduction in the number of years of life lost before age 85 in 2019 of 359,026, and that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in the number of hospital days in 2019 of 2.48 million. A rough estimate of the cost per life-year before age 85 gained in 2019 from drugs previously added to the PBS is $AUS 1388.

Highlights

• •We examine the relationship between long-run changes in the drugs provided by the Pharmaceutical Benefits Scheme (PBS) and mortality and hospital utilization in Australia, by analyzing the correlation across diseases between the change in the number of drugs used to treat the disease provided and the subsequent change in mortality or hospital utilization from that disease.
• •Diseases for which there were larger increases in the number of PBS drugs tended to have smaller subsequent growth in premature (before ages 85, 75, and 65) mortality and smaller growth in the number of hospital days 2–10 years later.
• •We estimate that the 1996–2013 increase in the number of PBS drugs was associated with a reduction in the number of years of life lost before age 85 in 2019 of 359,026, and that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in the number of hospital days in 2019 of 2.48 million.
• •A rough estimate of the cost per life-year before age 85 gained in 2019 from drugs previously added to the PBS is $AUS 1388.

Key points

• •Diseases for which there were larger increases in the number of Pharmaceutical Benefits Scheme (PBS) drugs tended to have smaller subsequent growth in premature (before ages 85, 75, and 65) mortality, and smaller growth in the number of hospital days 2–10 years later.
• •We estimate that the 1996–2013 increase in the number of PBS drugs was associated with a reduction in the number of years of life lost before age 85 in 2019 of 359,026, and that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in the number of hospital days in 2019 of 2.48 million.
• •A rough estimate of the cost per life-year before age 85 gained in 2019 from drugs previously added to the PBS is $AUS 1388.

1. Introduction

The Australian Government subsidizes the cost of many medicines for Australians through the Pharmaceutical Benefits Scheme (PBS). The PBS was established under Part VII of the National Health Act 1953 and is now regarded as a key component of the National Medicines Policy, which “aims to improve positive health outcomes for all Australians through their access to and wise use of medicines” (Storen et al. (2022)). In 2019, the government paid for 75% of Australia’s expenditure on prescribed medicines (OECD (2023a)).

As shown in Fig. 1, between 1992 and 2021, the number of drugs (WHO ATC5 chemical substances1) provided by the PBS increased by 49%, from 636 to 949.2 The objective of this study is to examine the relationship between long-run changes in the drugs provided by the PBS and both mortality and hospital utilization in Australia. Mortality reduction is an important component of economic growth, broadly defined. Nordhaus (2003) argued that “to a first approximation, the economic value of increases in longevity in the last hundred years is about as large as the value of measured growth in non-health goods and services.” Cutler et al. (2006) concluded that “knowledge, science, and technology are the keys to any coherent explanation” of mortality. Jones (1998, pp. 89–90) argued that “technological progress is driven by research and development (R&D) in the advanced world.” The pharmaceutical industry is one of the most research-intensive industries. In the U.S. in 2019, the pharmaceutical industry’s ratio of R&D to sales (16.3%) was 3.7 times as high as it was for all industries (National Science Board (2023)). In recent years, 88% of privately-funded biomedical research has been performed by pharmaceutical and biotechnology firms (Dorsey (2010)).

Fig. 1.

No. of WHO ATC5 drugs in PBS, 1992–2021.

To perform this assessment, we will analyze the correlation across diseases between the change in the number of drugs used to treat the disease provided and the change in mortality or hospital utilization from that disease. Fig. 2 shows the number of drugs used to treat four diseases (major causes of death) provided by the PBS during the period 1992–2021. The number of drugs used to treat diabetes increased by 25, from 17 to 42. The number of drugs used to treat acute myocardial infarction and ovarian cancer increased by smaller amounts, by 16 and 13, respectively. The number of drugs used to treat acute and subacute endocarditis declined, from 14 to 12. We hypothesize that the diseases for which there were larger increases in the number of PBS drugs had larger subsequent declines in mortality and hospitalization. To test this hypothesis, we will estimate two-way (by disease and year) fixed-effects (or “difference-in-differences”) models of the effect of the current or lagged number of PBS drugs on several measures of mortality and hospitalization.

Fig. 2.

No. of WHO ATC5 drugs in PBS for 4 diseases, 1992–2021.

Jones (2002) presented a model in which long-run growth is driven by the discovery (via research effort) of new ideas throughout the world.3 In general, measuring the number of ideas is challenging; some studies have used the number of patents as a proxy for the number of ideas. We believe that the principal innovation measure we will use—the change in the number of drugs available to treat a disease—is a useful, albeit imperfect, indicator of the increase in the number of ideas for treating the disease. A simple count of drugs might be a poor measure of medical effectiveness. However, poor measurement of an independent variable tends to bias the coefficient on that variable towards zero (Draper and Smith (1998), p. 31), which would make our estimates conservative.

We will also analyze an alternative measure of pharmaceutical innovation: the change in the number of drug classes (WHO ATC4 chemical subgroups) available to treat a disease. The addition of a new class of drugs might have a larger impact on mortality than the addition of another drug within the same class. However, the most important new drug classes may tend to include larger numbers of new drugs, so the number of drugs may be a better indicator than the number of drug classes. Also, the first drug in a class may not be the best drug in that class. Although atorvastatin was the fifth statin on the market, trailing lovastatin, the first, by more than 9 years, Schulze and Ringel (2013) refer to atorvastatin as a best-in-class late entrant that achieved major commercial success, capturing nearly double the peak annual sales of other statins.

We will also estimate some models that include an (imperfect) indicator of nonpharmaceutical medical innovation.

In the next section, we will describe the econometric models that we will estimate. In Section III, we will explain how the variables included in those models were constructed using data from a number of reliable sources. Estimates of the models will be presented in Section IV. Major implications of the estimates, including an assessment of the incremental cost-effectiveness of the drugs added to the PBS, will be discussed in Section V. The final section provides a summary and conclusions.

2. Econometric models of mortality and hospital utilization

2.1. Mortality model

Our estimates of the relationship between the (current or lagged) number of PBS drugs and mortality will be based on the following model:

ln(mort_dt) = β_k n_drug_d,t-k + α_d + δ_t + ε_dt (1)

where mort_dt is one of the following variables:

• yll85_dt = the number of years of life lost before age 85 due to disease d in year t (t = 2002, 2019)

• yll75_dt = the number of years of life lost before age 75 due to disease d in year t

• yll65_dt = the number of years of life lost before age 65 due to disease d in year t4

• age_death_dt = mean age at death from disease d in year t

and

• n_drug_d,t-k = ∑_s treat_sd pbs_s,t-k = the number of chemical substances to treat disease d provided by the PBS in year t-k (k = 0, 1, 2, …,10)

• treat_sd = 1 if chemical substance s is used to treat (indicated for) disease d5

• = 0 if chemical substance s is not used to treat (indicated for) disease d

• pbs_s,t-k = 1 if chemical substance s was provided by the PBS in year t-k

• = 0 if chemical substance s was not provided by the PBS in year t-k

• α_d = a fixed effect for disease d

• δ_t = a fixed effect for year t

Eq. (1) may be considered a health production function (Koç (2004)), and n_drug may be considered a measure of the stock of pharmaceutical “ideas.” Jones (2002) argued that “long-run growth is driven by the discovery of new ideas throughout the world.” The year fixed effects (δ_t‘s) in eq. (1) control for the effects of changes in macroeconomic variables (e.g. population size and age structure, GDP, educational attainment), to the extent that those variables have similar effects on mortality from different diseases.

Eq. (1) allows the relationship between the number of PBS drugs and mortality to be subject to a lag of up to 10 years. There is likely to be a lag between the initial inclusion of a drug in the PBS and its maximum association with mortality. Peak utilization of a drug tends to occur a number of years after it was first included in the PBS. The typical shape of the relationship between a drug’s “age” (number of years after initial PBS inclusion) and its utilization (number of “services” of the drug sold) is revealed by the age fixed effects (π_n‘s) of the following equation:

ln(n_services_sn) = π_n + γ_s + ε_sn (2)

where

• n_services_sn = the number of PBS services of drug s provided n years after initial PBS inclusion (n = 0, 1, 2, …, 15)

• π_n = a fixed effect for age n

• γ_s = a fixed effect for drug s

exp(π_n - π₅) is the mean ratio of utilization of a drug n years after initial PBS inclusion to its utilization 5 years after initial PBS inclusion. We estimated eq. (2) using annual utilization data for the period 1993–2016 on 558 drugs, weighting by total utilization of the drug during that period. The estimates of exp(π_n - π₅) are shown in Fig. 3. These estimates indicate that utilization of a drug tends to increase substantially during the first 6 years after initial PBS inclusion, increase more slowly during the next 5 years, and then begin to decline.

Fig. 3.

Estimates of the mean ratio of utilization of a drug n years after initial PBS inclusion to its utilization 5 years after initial PBS inclusion.

Some drugs may have to be utilized for several years to have their maximum association with mortality. On the other hand, drugs included more recently are likely to be of higher quality than older drugs; the effect of higher quality could offset the effect of lower quantity (utilization).

n_drug_d,t-k is the only disease-specific, time-varying regressor in eq. (1). As noted above, we will also estimate some models that include an (imperfect) indicator of nonpharmaceutical medical innovation. If the data were available, we would also control for disease incidence. Failure to control for incidence is unlikely to result in overestimation of the magnitude of β_k; exclusion of incidence may even result in underestimation of the magnitude of β_k. Higher disease incidence is likely to result in both higher disease burden and a larger number of PBS drugs:

Previous studies (Acemoglu and Linn (2004); Danzon et al. (2005)) have shown that both innovation (the number of drugs developed) and diffusion (the number of drugs launched in a country) depend on market size.

From eq. (1) (a model of the level of mortality), we can derive the following model of long-run mortality growth:

Δln(mort_d) = β_k Δn_drug_k_d + δ’ + ε’_d (3)

where

• Δln(mort_d) = ln(mort_d,2019) - ln(mort_d,2002) = the log-change from 2002 to 2019 in mortality from disease d

• Δn_drug_k_d = n_drug_d,2019-k - n_drug_d,2002-k = the change from 2002-k to 2019-k in the number of chemical substances to treat disease d provided by the PBS

• δ’ = δ₂₀₁₉ - δ₂₀₀₂

• ε’_d = ε_d,2019 - ε_d,2002

We will estimate eq. (3) by weighted least squares. For the first 3 measures of mortality, the weight will be (mort_d,2002 + mort_d,2019)/2. For the fourth measure of mortality (age_death_dt), the weight will be (n_deaths_d,2002 + n_deaths_d,2019)/2, where n_deaths_dt = the number of deaths caused by disease d in year t. The analysis will be performed on about 70 diseases included in the U.S. National Center for Health Statistics List of 113 Selected Causes of Death (New Jersey Department of Health (2023)). We exclude deaths from external causes.

The estimate of δ’ in eq. (3) is an estimate of mortality growth in the absence of expansion of the PBS, i.e., if mean(Δn_drug_k_d) = 0. This can be compared with mortality growth in the presence of expansion of the PBS: mean(Δln(mort_d)).

The long-run change in the number of chemical substances provided by the PBS (Δn_drug_k_d) might be correlated across diseases with other changes in disease treatment. Australian data on treatment methods, by disease and year, are not available. However, U.S. data on methods of treatment by office-based medical providers, by disease and year, are available from U.S. Medical Expenditure Panel Survey Office-Based Medical Provider Visits Event Files (Agency for Healthcare Research and Quality (2023)). The 1998 and 2015 files contain records of 104,740 and 172,388 patient visits, respectively. Each record indicates the patient’s diagnosis (coded by modified clinical classification code (CCS)), and whether the patient received each of the following: lab tests, a sonogram or ultrasound, x-rays, a mammogram, an mri/ctscan, an EKG or ECG, an EEG, a vaccination, anesthesia, other diagnostic tests or exams, or any of them. The fraction of visits in which at least one of those treatments was performed increased from 32% in 1998 to 40% in 2015. We computed the 1998–2015 change in the fraction of visits for each disease (152 CCS codes) in which any of those procedures were performed: Δany_ procedure%_CCS = any_ procedure%_CCS,2015 - any_ procedure%_CCS,1998, where any_ procedure%_CCS,t = the fraction of visits with patient diagnosis CCS in year t in which any procedure was performed. We also computed the 1998–2015 change in the number of PBS drugs using this disease classification: Δn_drug_CCS = n_drug_CCS,2015 – n_drug_CCS,1998. Then we estimated the following simple regression by weighted least-squares, weighting by the mean number of patient visits in 1998 and 2015: Δany_ procedure%_CCS = α + β Δn_drug_k_CCS + ε. The estimate of β (=−0.0039; t-value = −2.92; p-value = .0040) is negative and significant: diseases for which there were larger 1998–2015 increases in the number of PBS drugs had smaller 1998–2015 increases in the fraction of U.S. office-based medical provider visits in which any procedure was performed. This suggests that, to the extent that use of those procedures reduces mortality, failure to control for them may bias estimates of β_k in eq. (3) towards zero.

In addition to estimating eq. (3), we will estimate the following model, which includes an alternative measure of pharmaceutical innovation:

Δln(mort_d) = β_k Δn_class_k_d + δ’ + ε’_d (4)

where

• Δn_class_k_d = n_class_d,2019-k - n_class_d,2002-k = the change from 2002-k to 2019-k in the number of WHO ATC4 chemical subgroups to treat disease d provided by the PBS

We will also estimate a model that includes an (imperfect) indicator of nonpharmaceutical medical innovation as an additional explanatory variable. That measure is the (current or lagged) change in the mean vintage of Medical Subject Headings (MeSH) Analytical, Diagnostic and Therapeutic Techniques and Equipment (ADTTE) terms in thousands of PubMed articles about many diseases. PubMed is a database of 35 million journal articles published since 1946 in 5400 of the world’s leading biomedical journals (U.S. National Library of Medicine (2023a)). PubMED records are indexed with Medical Subject Headings (MeSH). MeSH is the U.S. National Library of Medicine’s controlled vocabulary thesaurus; MeSH consists of sets of terms (‘descriptors’) in a hierarchical structure that permits searching at various levels of specificity (U.S. National Library of Medicine (2023b)). The MeSH Section staff continually revise and update the MeSH vocabulary. Between 1955 and 2015, the number of MeSH descriptors increased from 15.8 thousand to 27.8 thousand. On average, about 200 descriptors were added per year. 8.6% of MeSH descriptors in 2016 were ADTTE descriptors.

The additional regressor included in eq. (3) is: Δtechnique_vintage_k_d = (technique_vintage_d,2019-k - technique_vintage_d,2002-k), where

• technique_vintage_k_d,t-k = (Σ_m first_year_m * freq_md,t-k)/(Σ_d freq_md,t-k)

• first_year_m = the year ADTTE descriptor m first appeared in PubMed

• freq_md,t-k = the number of times ADTTE descriptor m appeared in PubMed articles about disease d in year t-k (t = 2002, 2019; k= 0, 1, 2, …,10)

The equation that includes the indicator of nonpharmaceutical medical innovation as an additional explanatory variable is:

Δln(mort_d) = β_k Δn_drug_k_d + π_k Δtechnique_vintage_k_d + δ’ + ε’_d (5)

2.2. Hospital utilization model

Our estimates of the relationship between the (current or lagged) number of PBS drugs and hospital utilization will be based on the following model:

ln(hosp_dt) = π_k ln(n_drug_d,t-k) + α_d + δ_t + ε_dt (6)

where hosp_dt is one of the following variables:

• days_dt = the number of days of inpatient hospital care due to disease d in year t (t = 2002, 2019)

• discharges_dt = the number of hospital discharges due to disease d in year t

• alos_dt = average length of stay (in days) of hospital discharges due to disease d in year t

From eq. (6) (a model of the level of hospital utilization), we can derive the following model of long-run hospital utilization growth:

Δln(hosp_d) = π_k Δln(n_drug_k_d) + δ’ + ε’_d (7)

where

• Δln(hosp_d) = ln(hosp_d,2019) - ln(hosp_d,2002) = the log-change from 2002 to 2019 in hospital utilization for disease d

• Δln(n_drug_k_d) = ln(n_drug_d,2019-k) – ln(n_drug_d,2002-k) = the log-change from 2002-k to 2019-k in the number of chemical substances to treat disease d provided by the PBS

According to eq. (3), which is derived from the semi-logarithmic model of mortality, mortality growth is a function of the (absolute) change in the number of drugs. According to eq. (7), which is derived from the log-log model of hospital utilization, hospital utilization growth is a function of the log-change (approximately the percentage change) in the number of drugs. In the log-log model, an increase of n_drug from 1 to 2 (i.e., a doubling of n_drug) would have a larger impact than an increase of n_drug from 10 to 19 (i.e., less than doubling of n_drug). In principle, the log-log model could be either more or less appropriate than the semi-logarithmic model. We estimated both semi-logarithmic and log-log models of both mortality and hospital utilization. In the case of mortality, estimates of semi-logarithmic models were highly significant, and estimates of log-log models were insignificant. In the case of hospital utilization, estimates of log-log models were highly significant and estimates of semi-logarithmic models were insignificant. Hence, in section IV, we will present estimates of semi-logarithmic models of mortality (eq. (3)) and of log-log models of hospital utilization (eq. (7)).

Changes in hospital utilization may result from changes in government policies or payor pressure as well as from changes in PBS drug inclusions. We are not aware of any evidence that changes in government policies or payor pressure are correlated across diseases with changes in PBS drug inclusions, and especially with PBS changes years earlier. But absence of evidence is not the same as evidence of absence.

We will estimate eq. (7) by weighted least squares. For the first measure of hospital utilization, the weight will be (days_d,2002 + days_d,2019)/2. For the second and third measures, the weight will be (discharges_d,2002 + discharges_d,2019)/2. The analysis will be performed on about 79 diseases included in the International Shortlist for Hospital Morbidity Tabulation (OECD (2023b)).

3. Data sources

PBS drug data. Data on pbs_s,t-k (the drugs (WHO ATC5 chemical substances) provided by the PBS) during the period 1992–2015 were constructed from two sources: (1) PBS Item Reports (Services Australia (2023a)) for the years 1992–2015 provided annual data on the number of services of each item, and (2) the PBS item drug map (Department of Health and Aged Care (2023)) was used to determine the WHO ATC5 chemical substance corresponding to each item. Data on the drugs (WHO ATC5 chemical substances) provided by the PBS during the period 2015–2021 were obtained from (Services Australia (2023b)).

Data on the approved indications of each drug (treat_sd) were obtained from the Thériaque database, produced by the Centre National Hospitalier d’Information sur le Médicament (2023). This database contains information on over 30,000 drug products sold in France. For each product, the database provides (1) the WHO ATC code(s) of the substance(s) contained in the product, and (2) the ICD-10 codes of the product’s approved indications. This enabled us to compute the approved indications of each ATC code. Thériaque provided information about 87% (=811/928) of the 928 PBS drugs in 2019; those drugs accounted for 94% of PBS prescriptions.

Mortality data were obtained from the WHO Mortality Database (World Health Organization (2023b)).

Hospital utilization data were obtained from the OECD Health Statistics database (OECD (2023a)).

Appendix Table 1 contains a subset of the data used in the mortality analysis. Appendix Table 2 contains a subset of the data used in the hospital utilization analysis.

4. Estimates of models of mortality and hospital utilization

4.1. Mortality growth estimates

Estimates of β_k from the mortality growth model (eq. (3)) are presented in Table 1 and plotted in Fig. 4. Each estimate is from a separate regression. Panel A of the table and figure shows estimates when the dependent variable is Δln(yll85). When 0 ≤ k ≤ 3, the estimates of β_k are insignificant. When k > 3, the estimates of β_k are negative and at least marginally significant (p-value ≤.081), and the estimates of β₆, β₇, and β₁₀ are significant (p-value ≤.042). Diseases for which there were larger increases in the number of PBS drugs tended to have smaller growth in yll85 4–10 years later. The estimate of β₆ indicates that one additional drug for a disease was associated with a reduction in the number of years of life lost from the disease before age 85 of about 2%.

Table 1.

Estimates of β_k from mortality growth model (eq. (3)), Δln(mort_d) = β_k Δn_drug_k_d + δ’ + ε’_d

Note: each estimate is from a separate regression.

parameter	estimate	std. err.	t Value	Pr > |t|	parameter	estimate	std. err.	t Value	Pr > |t|
	A. dep. var. = Δln(yll85)					B. dep. var. = Δln(yll75)
β0	−0.008	0.008	−0.98	0.333	β0	−0.015	0.008	−1.85	0.068
β1	−0.006	0.008	−0.71	0.482	β1	−0.012	0.008	−1.54	0.127
β2	−0.011	0.008	−1.38	0.172	β2	−0.019	0.008	−2.41	0.019
β3	−0.010	0.008	−1.15	0.253	β3	−0.017	0.008	−1.99	0.051
β4	−0.016	0.009	−1.80	0.076	β4	−0.023	0.009	−2.52	0.014
β5	−0.017	0.009	−1.87	0.066	β5	−0.021	0.009	−2.27	0.026
β6	−0.020	0.009	−2.33	0.023	β6	−0.022	0.009	−2.57	0.012
β7	−0.018	0.009	−2.07	0.042	β7	−0.021	0.009	−2.52	0.014
β8	−0.015	0.009	−1.77	0.081	β8	−0.020	0.009	−2.27	0.026
β9	−0.017	0.009	−1.93	0.057	β9	−0.020	0.009	−2.33	0.023
β10	−0.021	0.009	−2.40	0.019	β10	−0.023	0.009	−2.67	0.009
	C. dep. var. = Δln(yll65)					D. dep. var. = Δage
β0	−0.019	0.008	−2.23	0.029	β0	0.069	0.029	2.35	0.022
β1	−0.016	0.009	−1.87	0.066	β1	0.063	0.029	2.20	0.031
β2	−0.023	0.008	−2.78	0.007	β2	0.080	0.031	2.57	0.012
β3	−0.020	0.009	−2.22	0.030	β3	0.069	0.031	2.23	0.029
β4	−0.024	0.009	−2.56	0.013	β4	0.070	0.034	2.05	0.045
β5	−0.021	0.010	−2.10	0.039	β5	0.038	0.035	1.10	0.277
β6	−0.021	0.009	−2.23	0.029	β6	0.033	0.032	1.02	0.312
β7	−0.021	0.009	−2.35	0.022	β7	0.052	0.032	1.62	0.110
β8	−0.020	0.009	−2.13	0.037	β8	0.057	0.032	1.77	0.081
β9	−0.020	0.009	−2.12	0.038	β9	0.050	0.032	1.59	0.117
β10	−0.022	0.009	−2.32	0.024	β10	0.048	0.032	1.50	0.138

Estimates in bold are statistically significant (p-value <.05).

Fig. 4.

Estimates of β_k from mortality growth model (eq. (3)), Δln(mort_d) = β_k Δn_drug_k_d + δ’ + ε’_d

Note: each estimate is from a separate regression.

Panel B of the table and figure shows estimates when the dependent variable is Δln(yll75). In this case, when k ≥ 2, 8 of the 9 estimates of β_k are negative and significant (p-value ≤.026). Fig. 5 shows a bubble plot of the correlation across diseases between the 1996–2013 change in the number of PBS drugs and the 2002–2019 change in ln(yll75). The bubble area is proportional to (yll75_d,2002 + yll75_d,2019)/2.

Fig. 5.

Correlation across diseases between 1996 and 2013 change in number of PBS drugs and 2002–2019 change in ln(yll75).

Panel C of Table 1 and Fig. 4 shows estimates when the dependent variable is Δln(yll65). In this case, 10 of the 11 estimates of β_k (including the estimate of β₀) are negative and significant. This suggests that the addition of drugs to the PBS was associated with reductions in mortality at lower ages sooner than it was at higher ages.

Panel D of the table and figure shows estimates when the dependent variable is Δage_death. The change in mean age at death is significantly positively related to the change in the number of PBS drugs 0–4 years earlier. It is most strongly related to the change in the number of PBS drugs 2 years earlier. The estimate of β₂ indicates that, on average, one additional drug for a disease was associated with an increase in mean age at death from that disease of about one month two years later.

As discussed above, we also estimated a model (eq. (4)) that includes an alternative measure of pharmaceutical innovation: the change from 2002-k to 2019-k in the number of WHO ATC4 chemical subgroups (“drug classes”) to treat disease d provided by the PBS (Δn_class_k_d). Only one of the 44 (4 mortality measures for k = 0, 1, 2, …,10) coefficients of Δn_class_k_d was statistically significant (p-value <.05). The change in the number of drugs is a much better predictor of mortality change than the change in the number of drug classes, perhaps because the most important new drug classes include larger numbers of new drugs.

We also estimated a model (eq. (5)) that includes an indicator of nonpharmaceutical medical innovation (Δtechnique_vintage_k_d) as an additional explanatory variable. None of the estimated coefficients on that variable were statistically significant. Perhaps Δtechnique_vintage_k_d is a poor measure of nonpharmaceutical medical innovation, which is more difficult to measure than pharmaceutical innovation. It may also be due to the relatively low R&D-intensity of physician and clinical services. Lichtenberg (2023) estimated that prescription drugs are about 12 times as R&D-intensive as other medical expenditure. He found that a measure of pharmaceutical innovation (mean drug vintage) had a significant negative impact on the disability of elderly Americans, but that a measure of other medical innovation (mean vintage of physician and clinical services) did not have a significant impact.

4.2. Hospital utilization growth estimates

Estimates of π_k from the hospital utilization growth model (eq. (7)) are presented in Table 2 and plotted in Fig. 6. Panel A of the table and figure shows estimates when the dependent variable is Δln(days). When k ≥ 2, the estimates of π_k are negative and significant (p-value ≤.024). Diseases for which there was larger growth in the number of PBS drugs tended to have smaller growth in the number of hospital days 2–10 years later. The estimates of π_k indicate that a 10% increase in the number of drugs for a disease was associated with a 3–4% reduction in the number of hospital days for the disease 2–10 years later.

Table 2.

Estimates of π_k from hospital utilization growth model (eq. (7)), Δln(hosp_d) = π_k Δln(n_drug_k_d) + δ’ + ε’_d

Note: each estimate is from a separate regression.

parameter	estimate	std. err.	t Value	Pr > |t|	parameter	estimate	std. err.	t Value	Pr > |t|
	A. dep. var. = Δln(days)					B. dep. var. = Δln(discharges)
π0	−0.266	0.165	−1.61	0.112	π0	−0.196	0.158	−1.24	0.218
π1	−0.293	0.162	−1.81	0.074	π1	−0.196	0.156	−1.25	0.213
π2	−0.413	0.149	−2.78	0.007	π2	−0.209	0.151	−1.38	0.172
π3	−0.303	0.132	−2.30	0.024	π3	−0.194	0.139	−1.39	0.169
π4	−0.348	0.135	−2.58	0.012	π4	−0.215	0.143	−1.50	0.138
π5	−0.395	0.135	−2.92	0.005	π5	−0.248	0.135	−1.83	0.071
π6	−0.350	0.129	−2.71	0.008	π6	−0.209	0.123	−1.70	0.094
π7	−0.364	0.129	−2.84	0.006	π7	−0.226	0.125	−1.81	0.075
π8	−0.369	0.117	−3.15	0.002	π8	−0.216	0.116	−1.86	0.067
π9	−0.348	0.117	−2.97	0.004	π9	−0.208	0.116	−1.79	0.078
π10	−0.338	0.113	−3.00	0.004	π10	−0.226	0.112	−2.01	0.048
	C. dep. var. = Δln(alos)
π0	−0.124	0.085	−1.46	0.149
π1	−0.137	0.084	−1.64	0.106
π2	−0.139	0.081	−1.71	0.091
π3	−0.111	0.075	−1.48	0.143
π4	−0.151	0.077	−1.97	0.052
π5	−0.137	0.073	−1.88	0.064
π6	−0.140	0.066	−2.12	0.037
π7	−0.143	0.067	−2.13	0.037
π8	−0.138	0.062	−2.21	0.030
π9	−0.142	0.062	−2.28	0.025
π10	−0.132	0.061	−2.18	0.033

Estimates in bold are statistically significant (p-value <.05).

Fig. 6.

Estimates of π_k from hospital utilization growth model (eq. (7)), Δln(hosp_d) = π_k Δln(n_drug_k_d) + δ’ + ε’_d

Note: each estimate is from a separate regression.

Panel B of the table and figure shows estimates when the dependent variable is Δln(discharges). In this case, only the estimate of π₁₀ is significant (p-value = .048), although all the estimates are at least marginally significant (p-value ≤.094) when 5 ≤ k ≤ 10.

Panel C of the table and figure shows estimates when the dependent variable is Δln(alos). In this case, the estimates are negative and significant (p-value ≤.037) when 6 ≤ k ≤ 10. This indicates that the larger the growth in the number of drugs for a disease, the greater the decline in average length of stay for the disease 6–10 years later. The reduction in the number of hospital days documented in Panel A appears to be primarily attributable to a reduction in average length of stay.

5. Discussion

As discussed above, our estimates allow us to compare the (actual) magnitudes of mortality and hospital utilization growth in the presence of previous PBS drug increases to estimates of the (counterfactual) magnitudes in the absence of previous PBS drug increases. Mortality growth comparisons are shown in Fig. 7. As shown in Panel A, the actual 2002–2019 decline in yll85 per 100,000 population was 29.0%. Our estimates imply that, if the number of PBS drugs had not increased between 1996 and 2013, the 2002–2019 decline in yll85 per 100,000 population would have been 7.6%, and yll85 in 2019 would have been 21.2% higher than it actually was. Actual yll85 in 2019 was about 1.69 million, so we estimate that the 1996–2013 increase in the number of PBS drugs was associated with a reduction in yll85 in 2019 of 359,026 (=21.2% * 1.69 million).

Fig. 7.

2002–2019 change in mortality: actual vs. estimated, in absence of lagged increase in no. of PBS drugs.

The estimates for yll75 and yll65 in Panels B and C of Fig. 7 indicate similar differences between actual and counterfactual mortality growth. The estimate in Panel B indicates that 88% (=1 – (−3.3%/-28.1%)) of the 2002–2019 decline in the yll75 rate was associated with the 1996–2013 increase in the number of PBS drugs. As shown in Panel D, we estimate that if the number of PBS drugs had not increased between 2000 and 2017, the 2002–2019 increase in mean age at death would have been reduced by 0.61 years, from 2.04 to 1.43 years.

Hospital utilization growth comparisons are shown in Fig. 8. As shown in Panel A, the number of hospital days per 100,000 population declined by 4.3% between 2002 and 2019. Our estimates imply that if the number of PBS drugs had not increased between 1994 and 2011, the number of hospital days per 100,000 population would have increased by 9.1% between 2002 and 2019,6 and that the number of hospital days in 2019 would have been 10.6% higher than it actually was. The actual number of hospital days in 2019 was about 23.38 million, so we estimate that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in the number of hospital days in 2019 of 2.48 million (=10.6% * 23.38 million). According to the OECD, aggregate expenditure on inpatient curative and rehabilitative care in 2019 was $AUS 56.2 billion, which implies that average expenditure per hospital day was $AUS 2405 (=$AUS 56.2 billion/23.38 million). Therefore, we estimate that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in hospital expenditure in 2019 of $AUS 5.97 billion (=2.48 million days * $AUS 2405 expenditure/day).

Fig. 8.

2002–2019 change in hospital utilization: actual vs. estimated, in absence of lagged increase in no. of PBS drugs.

As shown in Panel B of Fig. 8, between 2002 and 2019, average length of stay declined by 15.6%. Our estimates imply that if the number of PBS drugs had not increased between 1994 and 2011, average length of stay would have declined by only 13.0%.

Finally, we can compute a rough estimate of the cost per life-year before age 85 gained in 2019 from drugs previously added to the PBS. According to the OECD, between 2002 and 2019, total prescribed medicine expenditure increased from $AUS 7.41 billion to $AUS 16.11 billion, and the population increased from 19.50 million to 25.37 million, so per capita drug expenditure increased by $AUS 255, from $AUS 380 to $AUS 635. Suppose that the entire increase in per capita drug expenditure was due to the increase in the number of PBS drugs. (This is undoubtedly conservative, since part of the increase was due to an aging population.) Under that assumption, the increase in the number of PBS drugs increased aggregate drug expenditure by $AUS 6.47 billion (=25.37 million * $AUS 255), and the net increase in 2019 health expenditure associated with the lagged increase in the number of PBS drugs was $AUS 498 million (=$AUS 6.47 billion - $AUS 5.97 billion): the increase in drug expenditure minus the reduction in hospital expenditure. The cost per life-year before age 85 gained in 2019 was $AUS 1388 (=$AUS 498 million/359,026 life-years).

As noted by Bertram et al. (2016), authors writing on behalf of the WHO’s Choosing Interventions that are Cost–Effective project (WHO-CHOICE) suggested in 2005 that “interventions that avert one disability-adjusted life-year (DALY) for less than average per capita income for a given country or region are considered very cost–effective; interventions that cost less than three times average per capita income per DALY averted are still considered cost–effective.” In 2019, Australia’s per capita GDP was $AUS 78,092.

Although quality of life (QoL) in the additional life-years is likely to be less than perfect, the increase in DALYs is not necessarily less than the increase in life-years, since it is plausible that, in addition to delaying and preventing death, new drugs increased the QoL of people at a given number of years after diagnosis. A recent study of the U.S. (Lichtenberg (2022)) demonstrated that new drug class approvals reduced the number of people who were completely unable to work at a job, do housework, or go to school; the number of people with cognitive limitations; the number of people receiving Supplementary Security Income and Social Security; and the number of inpatient, outpatient, and home health visits.

6. Summary and conclusion

We assessed the relationship between long-run changes in the drugs provided by the PBS and both mortality and hospital utilization in Australia, by analyzing the correlation across diseases between the change in the number of drugs used to treat the disease provided and the subsequent change in mortality or hospital utilization from that disease.

Our estimates indicated that the diseases for which there were larger increases in the number of PBS drugs tended to have smaller subsequent growth in premature (before ages 85, 75, and 65) mortality. One additional drug for a disease was associated with a reduction in the number of years of life lost from the disease before age 85 of about 2%. The addition of drugs to the PBS was associated with reductions in mortality at lower ages sooner than it was at higher ages. The change in the number of drugs is a much better predictor of mortality change than the change in the number of drug classes, perhaps because the most important new drug classes include larger numbers of new drugs.

Diseases for which there was larger growth in the number of PBS drugs tended to have smaller growth in the number of hospital days 2–10 years later. The reduction in the number of hospital days appeared to be primarily attributable to a reduction in average length of stay.

We estimated that if the number of PBS drugs had not increased between 1996 and 2013, the number of years of life lost before age 85 in 2019 would have been 21.2% higher than it actually was, and that the 1996–2013 increase in the number of PBS drugs was associated with a reduction in the number of years of life lost before age 85 in 2019 of 359,026.

We also estimated that if the number of PBS drugs had not increased between 1994 and 2011, the number of hospital days in 2019 would have been 10.6% higher than it actually was, and that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in the number of hospital days in 2019 of 2.48 million. Average expenditure per hospital day was $AUS 2405 in 2019, so we estimate that the 1994–2011 increase in the number of PBS drugs was associated with a reduction in hospital expenditure in 2019 of $AUS 5.97 billion. A rough estimate of the cost per life-year before age 85 gained in 2019 from drugs previously added to the PBS is $AUS 1388. This figure is less than 2% of Australia’s per capita GDP.

The number of drugs provided by the PBS has increased, but it might have increased more than it did, and this might have been associated with larger reductions in mortality and hospital utilization. Table 3 shows data derived from the IQVIA MIDAS database on the number of post-2010 drugs (drugs first sold anywhere after 2010) sold in 2018 in 63 countries and regions. Australia ranks 26th on this list. The number of post-2010 drugs sold in Australia was about half as large the number sold in the three top countries (the USA, Germany, and the UK), and was lower than the number sold in Russia, Hungary, and Mexico. Several previous studies (Lichtenberg (2019a, 2019b, 2023)) have shown that countries or states with greater access to new drugs had lower mortality, disability, and hospitalization, controlling for unobserved factors such as the average quality of medical care in a country or state.

Table 3.

Number of post-2010 drugs sold in 2018, by country or region.

rank	country or region	Number of post-2010 drugs sold in 2018	rank	country or region	Number of post-2010 drugs sold in 2018
1	USA	181	33	SAUDI ARABIA	68
2	GERMANY	160	34	UAE	65
3	UK	160	35	TURKEY	64
4	AUSTRIA	147	36	MALAYSIA	63
5	ITALY	141	37	ROMANIA	60
6	SWEDEN	138	38	BULGARIA	59
7	NORWAY	133	39	LEBANON	59
8	SPAIN	127	40	INDIA	52
9	FRANCE	124	41	LUXEMBOURG	45
10	FINLAND	123	42	EGYPT	44
11	PORTUGAL	120	43	NEW ZEALAND	44
12	SWITZERLAND	120	44	CENTRAL AMERICA	43
13	CANADA	112	45	CHINA	43
14	BELGIUM	111	46	GREECE	43
15	JAPAN	106	47	CHILE	41
16	PUERTO RICO	102	48	SERBIA	41
17	KOREA	100	49	ECUADOR	40
18	SLOVENIA	98	50	PHILIPPINES	40
19	SLOVAKIA	96	51	ARGENTINA	36
20	RUSSIA	93	52	SOUTH AFR	33
21	HUNGARY	92	53	PERU	29
22	IRELAND	89	54	COLOMBIA	28
23	MEXICO	85	55	BOSNIA	26
24	POLAND	85	56	INDONESIA	25
25	HONG KONG	84	57	BELARUS	24
26	AUSTRALIA	82	58	KUWAIT	16
27	BRAZIL	82	59	PAKISTAN	15
28	TAIWAN	82	60	FRENCH WEST AFRICA	13
29	CROATIA	79	61	MOROCCO	13
30	SINGAPORE	76	62	TUNISIA	9
31	CZECH REPUBLIC	74	63	ALGERIA	7
32	THAILAND	70

Source: Author's calculations based on IQVIA MIDAS database.

Compliance with ethical standards

Informed consent and ethics approval were not required.

Financial support for this research was provided by Medicines Australia.

Author conflicts of interest: none.

Declaration of competing interest

The author declares the following financial interests/personal relationships which may be considered as potential competing interests: Frank R. Lichtenberg reports financial support was provided by Medicines Australia. Frank R. Lichtenberg reports a relationship with Pfizer Inc that includes: funding grants. Research grant from Novartis.

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Data availability

Data will be made available on request.