Market watch: Forecasting market share in the US pharmaceutical market

Drug development is costly so drug makers need accurate estimates of sales potential. However, sales forecasts are often unreliable. Our study is unique in combining a large sample of drug classes with data on entry order and promotional spending to estimate peak market share while controlling for product quality. We estimate peak market share of 50 drugs (covering 29 different therapeutic classes) from the past two decades based on the promotional spending, entry order, speed-to-market and subsequent market entry of additional competitors. We control for product quality differences by considering only drugs that regulators classified as clinically comparable to other drugs in the class. We find that a product’s peak market share will be about 32% if: i) it is second to the market, ii) it is 2-years delayed relative to the first entrant (of comparable quality to the first entrant), iii) it receives as much promotional spending as the first entrant, and iv) it is followed by a third entrant 1 year later. The results can help managers more accurately forecast sales, and illustrate the value of earlier product launch.


Introduction
Drug development is costly, 1,2 so drug makers need accurate estimates of sales potential. However, sales forecasts are often unreliable. 3 Here, we present an analysis of data concerning entry order and promotional spending from a large sample of drug classes, to estimate peak market share while controlling for product quality.

Data and Methods
The data sample included new molecular entities approved by the US Food and Drug Administration (FDA) from 1988 to 2009. We ended the sample at 2009 approvals so that each assessed product would accrue at least 4 years of post-launch data on sales and promotion. To control for product quality, we focused on drugs which the FDA granted standard review (meaning that the drugs did not represent a significant advance over previous drugs) and that the French Transparency Commission classified as providing little or no improvement over previous drugs. 4 The sales data came from the IMS Health US National Prescription Audit from IMS Health, and the promotional data came from SDI Health.
We used an ordinary least-squares regression. The dependent variable was peak share which we defined as the maximum monthly share reached by a new entrant during the first 4 years on the market. The independent variables were share of promotional spending, 5,6 order of entry, 7,8 time-tomarket, and whether new competitors entered for second entrants. We calculated the share of promotional spending as the ratio of the new entrant's promotional spending to the total promotional spending from all products in the therapeutic area during the first 12 months post-launch, where promotional spending included physician/nurse detailing, journals, events and direct-to-consumer advertising. We measured time-to-market in quarters relative to the most recent entrant on the market.
We included indicator variables for third and fourth entrants (named third and fourth respectively), so for second entrants, third=fourth=0. The variable new_competitor equaled 1 if a second entrant faced a third entrant.
For more details about the methods, identification strategy, and alternative specifications, please see the appendix.

Results
Our sample comprised 29 second entrants, 13 third entrants and 8 fourth entrants. We estimated peak market share as follows: Order of entry. Relative to a second entrant, peak market share was 18 percentage points lower for a third entrant and 23 percentage points lower for a fourth entrant, even if they had the same promotional spending.
Time-to-market. For a second entrant, each additional delay of one quarter led to a peak market share decrease of 0.9 percentage points. The impact of a delay for a third or fourth entrant was smaller.
New competitor. The launch of a third entrant reduced the peak market share potential of second entrants by 6 percentage points. Entry of a fourth competitor did not have a statistically significant effect on the third entrant.
Given estimates of the value of reduced time to market, we can also calculate the value of a priority review voucher, which decreases FDA review time from about 10 months to 6 months. 9 Previous estimates of voucher value were based on the value gained by shifting existing sales earlier in time; however, we show that having an earlier launch also increases peak market share. For example, if the second entrant reached the market 4 months earlier, peak market share would increase by 1.2 percentage points, or US$12 million in the peak year for a US$1 billion drug (in addition to the value of shifting sales earlier). Figure 1 summarizes the determinants of peak share. The promotional share assumptions used were the average share in our sample (i.e. 53% for a second entrant, 29% for a third entrant and 24% for a fourth entrant, see Table 1   Forecasting a drug's peak market share is challenging. We hope that the model presented in this paper will give managers additional insights about the future success of investigational drugs.

Disclosures
Stephane A. Régnier is an employee of Novartis Pharma AG. The views in the article are those of the authors, not necessarily those of Novartis Pharma AG.

Data
The sample included new molecular entities approved by the US Food and Drug Administration (FDA) from 1988 to 2009. We ended the sample at 2009 approvals so that we would have four years of sales and promotion following launch. We identified drugs for the sample using the universe of FDA approvals between 1999 and 2009 1 and annual reports from the 20 leading pharmaceutical companies from 2007 to 2009. We excluded drugs primarily used in hospitals and clinics, because our sales data were collected from retail pharmacies. We also excluded priority drugs, focusing on drugs receiving standard FDA review (meaning that they did not represent significant advances over previous drugs) and received an Amélioration du Service Médical Rendu rating of IV (minor improvement in terms of efficacy or safety) or V (no improvement) by the French Transparency Commission. 2 Finally, we excluded drugs in classes with generic competition. The list of drugs included in the analysis appears in Table S1.

Missing data
If total prescriptions in the fourth year were not available (e.g. Allegra We did not have data from the amélioration du Service Médical Rendu (ASMR) ratings for Pravachol (pravastatin), Astelin (azelastine), Detrol+LA (tolterodine) and Enablex (darifenacin). We used our judgment to classify those products as me-too. We also used our judgment to classify Actonel (risedronate) as a me-too even though the product received an ASMR of III (modest improvement).

Combinations
The sales and promotional spending of Cialis (tadalafil) (

Co-promotions
When products were co-promoted by two pharmaceutical companies, the higher revenue of the two firms was used to calculate ratio of total company revenue to total revenue of competing companies.
The list of identified co-promotions is available on request.

Me-too classification
Three products (Actos Five products were excluded even if they received an ASMR of IV or V due to unusual circumstances (such as inferior delivery device, multiple drugs sold by the same company in a therapeutic class, etc.). Including those eight products in the model reduced the precision of the model (R 2 of 0.74 instead of 0.89) but did not change the conclusions. The analysis including the excluded products is available on request. Three products (Pylera, Tyzeka and Uroxatral) were not included because their ASMR ratings were not retrieved.

Methods
To estimate the impact of the share of promotional spending, order of entry and speed-to-market on peak market share, we used an ordinary least-squares regression.
In the econometric model, an observation is a drug.

Identification
One major methodological concern with estimating the effect of promotional spending and other factors is that they may be correlated with unobserved characteristics of the product. For example, if the drug is high quality relative to competitors, and the manufacturer promotes the drug more as a result, then we will overestimate the effect of promotional spending on market share, wrongly assuming that the high market share was due to promotion rather than due to quality.
Fortunately, the nature of the regulatory process allows us to eliminate quality extremes. Products that are much lower quality than competitors are not approved by regulators. Products that are much higher quality than competitors (as measured by US and French regulators) were not included in our sample.
A further methodological challenge was accounting for the direction of causality; that is, whether sales depend on promotion, or promotion depends on expected sales. 4 This potential issue was again mitigated by considering only drugs that had no significant improvements over previous drugs. For those drugs, it is unlikely that manufacturers have private information and forecast the success/failures of drugs and set promotional spending at launch accordingly.
To further address concerns about causality, we used instrumental variables that affect promotional spending but not product quality or market share. We conducted a two-stage least-squares regression followed by an endogeneity test (endogtest in ivreg2 command in Stata based on Sargan-Hansen statistics). In the first stage, we hypothesized that larger companies have more resources and, as a consequence, have a larger share of promotional expenses than smaller firms. We also assumed that a US location of company headquarters for new entrants and first entrants could potentially impact the level of promotional spending. 5 We assumed that there was no direct effect between firm size and peak share (other than through higher promotional expenses). The endogeneity test described in the next section did not find evidence of endogeneity.

Finite sample assumptions validation
The model relies on least-square estimates with a small number of observations, so we used the following tests to ensure ordinary least-square assumptions for finite samples were not violated: (1)

Results
Order of entry, time-to-market, promotional spending, and launch of additional competitors were all statistically significant variables (Table S3). The impact of each variable is described in the article. Figure 1 in the main article illustrates the impact on market share of promotional share, order of entry, and the launch of a third entrant. Figure 1b summarizes the determinants of peak share. The promotional share assumptions used were the average share in our sample (that is, 53% for a second entrant, 29% for a third entrant and 24% for a fourth entrant, see Table S2). For a second entrant (the left panel of Figure 1b), peak share was 34%, assuming 53% promotional share, a 2-year delay in reaching the market, and a new entrant later. For a third entrant (the middle panel of Figure 1b), the peak share was 17 percent, because it was a later entrant and had less promotional spending. The peak share was 12 percent for a fourth entrant (the right panel of Figure 1b). The impact of a delay by order of entry is shown in Figure S1. A 6-month delay is associated with a peak share reduction of 1.8percentage point for a second entrant (vs. 0.4 point for a third entrant)

Identification
The results from the two-stage least-squares regression appear in Table S4. The coefficients and significance levels are similar to those in the base case. It was not possible to reject that the share of promotional spending was exogenous (p = 0.60). Therefore, there is limited risk that our model suffers from endogeneity issues.
We ran several tests to assess the validity of the instruments. The Hansen J statistic was 1.63 (p = 0.44). Therefore, we cannot reject the hypothesis that the instruments are valid instruments, i.e.
uncorrelated with the error term. In addition, the instruments were not weak (partial R 2 of excluded instruments of 0.33 with an F-value of 8.6).

Finite sample assumptions validation
The ordinary least-square finite sample assumptions do not appear to be violated based on the results from the Link-RESET, Breusch-Pagan, Park and Shapiro-Wilk W tests (p = 0.96, p = 0.16, p = 0.10 and p = 0.14 respectively). The highest VIF value was 4.8 and we found no patterns between residuals and predicted values.

Accuracy and generalizability of the results
For 38 products, the values of predicted and actual peak share differed by fewer than 5 percentage points ( Figure S2). The median error (in absolute terms) in prediction was 2.9 percentage points (median actual share is 25.0%). Expressed as a percentage of sales, the median error of prediction was 14.7%. Details below also show that the ordinary least squares regression assumptions were not violated. The strength of the model resides in the comprehensiveness of the variables explaining share. For instance, the model explains much of the variance for peak shares (adjusted R 2 is 0.87).
Therefore, most, if not all, relevant variables are included in the model.
We also investigated whether the results could be generalized outside the selected sample, because the number of explanatory variables was large relative to the number of observations. If the model is too We conclude that the evidence for having a too-complex model was not strong and, therefore, the results should be valid outside the sample to make an average forecast prediction.