Response to Lichtenberg

Dear Editor,

Grootendorst, Piérard and Shim (GPS) critically appraise the literature examining the impact of pharmaceutical use and innovation on life expectancy (LE) [1]. They point to methodological weaknesses in the studies and find that the outputs of the models are not robust to seemingly minor changes in specification. Lichtenberg objects to the GPS appraisal. He makes a series of claims:

• • Claim 1: GPS do not clearly distinguish between the effect of pharmaceutical expenditure on LE and the effect of pharmaceutical innovation on LE;

• • Claim 2: GPS overstate the degree of consensus about the value of pharmaceuticals (or pharmaceutical innovation);

• • Claim 3: GPS fail to completely appraise the literature;

• • Claim 4: GPS claim that “policy-makers typically are concerned with the impact of individual drugs (or perhaps groups of therapeutically similar drugs) on health outcomes,” but some important pharmaceutical policy questions depend on the impact of drugs in general;

• • Claim 5: GPS make conclusions that are less robust to their sensitivity analysis than some key findings from previous studies;

• • Claim 6: GPS misinterpret estimates of the parameters of this model;

• • Claim 7: GPS do not provide any evidence to support their claim that “analyses of mortality using individual-level data or aggregated disease-specific data are more promising,” or acknowledge the limitations of those analyses.

Our view is that, while Claim 6 has merit, the remaining points do not. In the following sections, we explain why.

Claim 1

This claim is simply incorrect. Contrary to Lichtenberg’s assertion, we did indeed distinguish the respective effects of pharmaceutical expenditure and pharmaceutical innovation on LE (see page 354) [1]. Moreover, on the same page, we note that Lichtenberg’s study is exceptional in that it attempts to isolate the effect of pharmaceutical innovation on LE.

Claim 2

Grootendorst, Piérard and Shim state: “there is no question that pharmaceutical drugs are among the most cost effective forms of healthcare, and hold the promise to treat more effectively various debilitating health problems”. Lichtenberg says that this statement mischaracterizes consensus estimates of the productivity of pharmaceuticals. He points to several commentators who note that most new drugs are not therapeutic breakthroughs but, at best, provide incremental improvements over existing drugs.

Lichtenberg’s criticism is unfounded. These commentators speak merely of heterogeneity in the effectiveness of new drugs. They do not imply that, on average, new drugs are ineffective. We suspect that even the “innovation deniers” he refers to (at least one of whom prescribes drugs routinely) would agree that some new pharmaceutical drugs are therapeutically invaluable and are more cost effective than other medical interventions. For instance, we suspect that none of these commentators would claim that β-blockers are ineffective for the management of cardiovascular disease. However, some might claim that the ninth β-blocker on the market adds only marginal therapeutic value. Lichtenberg, in a different study, finds evidence of heterogeneity in the productivity of new drugs [2]. Specifically, he found that the launch of ‘priority review’ drugs decreased the US’s cause-specific mortality, whereas the launch of ‘standard review’ drugs (which constitute the majority of drug launches) did not.

Claim 3

Lichtenberg claims that our review is incomplete because we fail to appraise several studies that focus on the impact of pharmaceuticals use and innovation using disaggregated data. This point is puzzling. Our title and introduction make it clear that we focus on studies of LE. The LE measure used in these studies summarizes into a single number the age-specific mortality rates experienced in a population during a given year. The other studies that Lichtenberg refers to focus on mortality in subpopulations of individuals with specific conditions or those using specific drugs. We suggest that these are exactly the types of studies that are likely to yield better estimates of pharmaceutical productivity.

Claim 4

Lichtenberg takes issue with the statement “policy-makers typically are concerned with the impact of individual drugs (or perhaps groups of therapeutically similar drugs) on health outcomes”. He cites the example of US policy-makers who are charged with the task of choosing the number of years of market exclusivity for biologic drugs. These policy-makers, he seems to imply, seek the “theoretically optimal” period of exclusivity, which depends on, inter alia, “the magnitude of the health benefit … from the average new biotech drug.”

While we are mildly surprised that US policy makers are so technically sophisticated, we agree that there may be situations in which policy-makers could use information on the value of new drugs generally. Our point was that research into the effectiveness of individual drugs or therapeutically related groups of drugs is likely to be most useful to policy-makers. Indeed, the Obama administration has recently pledged US$1.1 billion “to compare drugs, medical devices, surgery and other ways of treating specific conditions” [101].

We also note that it may be possible to derive an estimate of the value of pharmaceutical innovation generally by combining estimates of the value of new drugs in managing cancer, heart disease and other high-prevalence diseases that are responsible for the bulk of premature mortality.

Claim 5

Lichtenberg takes issue with our respecification of his LE model. We argued that drug innovation, measured using the number of new molecular entities (NMEs) launched each year, affects LE via its effect on drug spending. On the basis of this logic, we estimated the ‘reduced form’ model; this model excludes the drug spending covariate and allows us to estimate the net effect of pharmaceutical innovation on LE. Lichtenberg argues that it is the combination of NMEs and drug spending that affects LE; in other words, each of these exert an independent effect on LE. We accept this criticism but found it surprising that Lichtenberg does not heed his own advice – he fails to include pharmaceutical spending as a covariate in any of the five models he presents. In his preferred model specification (his model three), he excludes private drug spending, which accounted for the majority of total drug spending in the USA over his sample period (according to the Organisation for Economic Co-operation and Development health database, the private share of prescription drug spending ranged from ∼97% in 1960 to 76% in 2001). He does include a covariate representing total publicly financed healthcare spending, but this covariate restricts the impact of pharmaceutical spending on LE to be the same as the impact of nonpharmaceutical spending on LE – a restriction he does not formally test. Moreover, in related work, he also fails to include pharmaceutical spending as a covariate alongside new drug launches [2].

Nevertheless, we followed Lichtenberg’s advice and re-estimated our model of US LE at birth, using the same dataset as before but also including the one period lag of log real per capita (RPC) pharmaceutical spending. Using these estimates, we found the steady-state elasticities of LE with respect to NMEs, as well as the LE elasticities with respect to RPC public-health, hospital and physician expenditure. The results from this revised model are a mixed blessing for Lichtenberg’s argument(Table 1). In this revised model, public health spending had no significant impact on LE, and NMEs and drug spending were highly effective. However, physician services spending had a pronounced, deleterious effect on LE. Hence, the estimated model confirms his hypotheses regarding pharmaceutical innovation but suggests that an even better policy would be to cut spending on physicians’ services. How can readers have confidence in the estimates that are favorable to his hypothesis when other estimates from the same model suggest some misspecification?

Why are LE regressions likely to be mispecified? We reiterate the point made in GPS: it is difficult to estimate a parsimonious model of LE that can isolate the impact of pharmaceutical innovation from surgical, diagnostic, medical device and disease management innovation, not to mention improvements in public health; the dissemination of information on health risks (which has resulted in changes in tobacco use, diet and other health-related behaviors); improvements in transport, occupational and product safety; and changes in the socioeconomic and demographic conditions that impact on LE. The difficulties of identification are compounded by the facts that, first, these factors operate on LE with different latency periods and, second, pharmaceutical innovation can be complementary to medical innovation (i.e., advances in biomedical knowledge allow drugs to be used more effectively). For instance, the Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial (ALLHAT) study on the efficacy of different antihypertensive drugs [102], and the Woman’s Health Initiative study of postmenopausal hormone therapy [103], have improved our understanding of appropriate pharmacotherapy in these disease areas.

On page 376 of his paper, Lichtenberg admits that pharmaceutical innovation is but one of several factors that could be responsible for the gains in LE observed since 1960, yet he fails to recognize this fact when interpreting his parameter estimates [3]. His rejoinder to our article is silent on this pivotal issue.

Claim 6

Lichtenberg correctly notes that we misinterpreted his estimates of the “bang for buck” spent on NMEs and other inputs. We thank him for pointing out this error.

Both Lichtenberg and GPS estimated the percentage change in LE at birth owing to a 1% increase in each of the inputs. However, we used these elasticity estimates in different ways. Lichtenberg used the estimate to compute the absolute change in LE at birth due to a 1-unit change in the input. He then converted the absolute change in LE at birth into the number of life-years saved in the US population (by multiplying by the mean annual number of births in the USA). He also converted the cost of a 1-unit change in the input into the total cost for the US population (e.g., the total cost of a US$1 increase in per capita health spending is $1 × the size of the US population).

Grootendorst, Piérard and Shim used the elasticity estimates to work out the change in the input required to increase LE at birth by 1 year (see equation (4) from GPS). We converted each of the requisite changes in the inputs into per capita terms. For the input RPC public health spending, no conversion was necessary. To find the RPC cost of the requisite increase in NMEs, we used the same average NME development cost estimate used by Lichtenberg (US$500 million) and divided this by the US population size over the sample period (225 million).

We have since discovered an error in GPS. Owing to a programming error, we miscalculated the increase in RPC public-health spending needed to increase LE at birth by 1 year. We reported in GPS that only US$0.03 was required. The corrected estimate is much higher: US$5.56 (95% CI: 2.76–8.36). The corrected estimate, however, is still less than the estimated RPC cost of increasing LE by conducting pharmaceutical research and develoment (US$30.66; 95% CI: 9.77–51.54).

Claim 7

Grootendorst, Piérard and Shim conclude that analyses using individual-level data or perhaps disease-specific data will probably produce more compelling results than LE regressions. Lichtenberg argues that:

“…since they have only performed sensitivity analyses of studies based on aggregate data, they have not established the superiority of studies based on patient-level or disease-specific data. Their conclusion is really a conjecture.”

He then argues that studies of disaggregated data present their own problems. We are not convinced by Lichtenberg’s counterarguments.

While we did not perform any sensitivity analyses of studies based on disaggregated data, these data have compelling theoretical reasons to recommend them. The most important of these is that disaggregated data offer more ways to control for confounding influences. Recent advances in econometrics, including instrumental variables techniques, difference in differences and regression discontinuity designs [4], enable better exploitation of the data. Moreover, studies that have focused on technological innovation in the management of specific conditions – including heart attacks, cataracts, low-birthweight infants, breast cancer, depression and high blood pressure – have exploited knowledge of the evolution of treatment patterns in these areas to refine estimates of pharmaceutical productivity [5,6].

How serious are the limitations of disaggregated data enumerated by Lichtenberg? They can be challenging, but not to the same degree as the challenges involved in analyzing aggregate LE data. He argues:

“Individual-level studies that involve large numbers of drugs (hence enabling inferences about the general impact of pharmaceutical innovation) are unlikely to be based on samples in which drugs are randomly assigned to patients.”

We agree; however, nonrandom assignment is a hallmark of virtually all studies based on observational data. In addition, this is not an insurmountable problem. Lichtenberg himself has made substantive contributions to the literature using exactly these types of data. He continues:

“Data on potentially confounding variables (e.g., education, race and income) are frequently unavailable at the patient level, and the follow-up period may be short.”

Our response is that the econometric techniques reviewed by Angrist and Pischke [4] can sometimes be used (depending on the context) to resolve issues of confounding bias when data on potential confounding variables are unavailable. We concede that follow-up periods can be insufficiently short to track the health outcomes of drug use (rendering any econometric approach useless). However, as the hypertension study conducted by Cutler et al. indicates, alternative approaches can be used instead [5]. Finally, Lichtenberg argues that disease-specific studies are subject to competing risks bias; for example:

“… using a newer drug to treat cardiovascular disease might reduce cardiovascular disease mortality but increase life-years lost due to cancer”.

Why can this problem not be remedied by focusing on all-cause mortality of those with cardiovascular disease?

Our view remains that LE regressions are not sufficiently informative to disentangle the impact of pharmaceutical innovation from other forms of medical innovation and other confounding factors. Lichtenberg does not contest this point. Disaggregated data, while not a panacea, offer credible ways of addressing this question.

Yours faithfully,

Table 1. Revised estimates of cost to increase life expectancy at birth by 1 year.

Input to affect life expectancy at birth	Per capita cost to increase life expectancy at birth by 1 year (US$)	95% confidence interval (US$)
		Lower limit	Upper limit
New molecular entity	21.92	13.72	30.12
Public health	25.91	-28.72	80.53
Physician services	-83.24	-156.10	-10.39
Hospital services	85.45	25.03	145.87