Direct and indirect savings from parallel imports in Sweden.

BACKGROUND: The aim was: i) to quantify the direct and indirect savings from parallel imports in Sweden during a period when sellers were forbidden from giving discounts to pharmacies, and ii) to study if the effects of competition from parallel imports on list prices became smaller in absolute size when sellers were allowed to give discounts to pharmacies.
METHODS: We analyzed the monthly prices for 3068 products during 61 months when discounts were forbidden and for 2504 products during 84 months when discounts were allowed. The price effects were estimated using dynamic models that rendered lagged numbers of competitors into valid and strong instruments for the current values.
RESULTS: When discounts were forbidden, parallel imports had a market share of 16% and were on average 9% cheaper than locally sourced drugs, which yielded a direct saving of 231 million Swedish kronor (SEK) (24 million EUR) per year. Also, parallel imports reduced the prices of products with the same substance by, on average, 6% in the long-term, which yielded indirect savings of 421 million SEK (44 million EUR) per year. In total, parallel imports reduced the cost for on-patent pharmaceuticals by 4%. When discounts were allowed, the average gap in list price between parallel imports and locally sourced products was reduced to 0.8%, and the list prices of locally sourced products were no longer significantly affected by competition from parallel imports.
CONCLUSION: When discounts were allowed, the savings of parallel imports through lower list prices were replaced by savings of pharmacies through secret discounts.

Background

In an attempt to practice third-degree price discrimination, producers may charge wholesalers in low-income countries less than they charge wholesalers in high-income countries. Parallel traders take advantage of these price differences by buying products intended for low-price countries and, without authorization from the patent holder, selling them to wholesalers in high-price countries. Parallel trade is allowed within the European Economic Area to help fulfill the objective of creating a single market.

This paper evaluates the savings from parallel imports in Sweden during a period when sellers of pharmaceuticals were forbidden from giving discounts to pharmacies. This situation implied that the official list prices were actual transaction prices, which enabled us to quantify the total savings. We also studied whether the effects of competition from parallel imports on list prices became smaller in absolute size when sellers were allowed to give discounts to pharmacies. There are several reasons why firms might prefer to give discounts, rather that lower official list prices, to reduce the market share of parallel imports. Lower list prices, for example, can reduce the revenues from other countries where the Swedish list prices are used as external reference prices,1 and, in Sweden, list prices might not be allowed to be increased within the benefit scheme when the competition diminishes.

With access to data on market shares for parallel imports and relative prices, direct savings can be easily calculated. This has been done previously, for example, by West and Mahon [2], who reported direct savings for Sweden of 424 million Swedish kronor (SEK) in 2002 (measured in retail prices).

West and Mahon [2] also showed price plots and comparisons of price changes over 5–6 years, which indicated that parallel imports exerted downward pressure on prices. Estimating this effect is difficult, however, because parallel imports are more likely to be sold the higher the prices of the locally sourced products are, rendering the variable endogenous. To address this endogeneity problem, Ganslandt and Maskus [3] and Granlund and Köksal-Ayhan [4, 5] used exchange rates and the age of drugs as instruments for competition from parallel imports and reported point estimates suggesting that competition from parallel imports reduced the prices of locally sourced drugs in Sweden by 12 to 21%. However, these instruments may affect the prices of locally sourced drugs in other ways than through the existence of parallel imports, which can create bias. For example, with a stronger Swedish currency, a producer can reduce the nominal price in Sweden without having to reduce the price in countries where the maximum allowed prices depend on Swedish prices measured in Euros.

Vandoros and Kanavos [6] instead used instruments based on the number of policies promoting parallel imports and the distance between the source countries and the four destination countries they analyzed (Germany, Sweden, the Netherlands, and the United Kingdom). They found no statistically significant price effect, but because of large standard errors, they could neither reject the premise that the price effect was large.2 Vandoros and Kanavos also analyzed the effect of the market share of parallel imports—following Kanavos and Costa-Font [7] and Kanavos and Vandoros [8]—but, like the previous studies, they found no statistically significant price effects. Christian Gollier, the discussant to Kanavos and Costa-Font, mentioned as one potential explanation for the lack of a significant estimate the possibility that “the local manufacturer actually matches the price of the importers by using hidden discounts to distributors rather than reducing the list price” [6 , p. 793].

To overcome the problem with weak and potentially endogenous instruments, Granlund [9] used a dynamic model that allowed lags of competition variables to be used as instruments for their current values. This approach yielded sufficiently many strong instruments for also studying the causal effects on the intensive margins, i.e., how the number of parallel traders and the number of therapeutic competitors affect prices. Granlund [9] used part of the data used in the present study: that for tablets and capsules sold in October 2002–October 2007. For this study, we estimated similar price functions as Granlund [9], but also did so for the period of January 2011–December 2017, for all forms of administrations, and calculated the direct and indirect savings yielded by parallel imports.

Rules regarding parallel imports

All Swedish residents are covered by a mandatory and uniform pharmaceutical benefit scheme. Since October 2002, a substitution legislation requires that pharmacy personnel inform consumers if cheaper substitute products are available, unless the prescriber has vetoed substitution or if the pharmacist has reasons to believe that the patient would be adversely affected, e.g., when the low-cost alternative has a package that the patient would find difficult to open. The Swedish Medical Products Agency defines a product as a substitute if it has the same active substance, strength, and form of administration (e.g., pills or oral fluid) and nearly identical package size.3 If consumers oppose the substitution, or choose to switch to a substitute other than the cheapest one available, they will be charged the entire incremental cost. For parallel imports, available substitutes are defined as those in stock at the pharmacy in question [10].4

Pharmaceutical producers and parallel traders are free to set their own prices, but to be included in the pharmaceutical benefits scheme, they must submit their prices for month t to the Pharmaceutical Benefits Agency (PBA) in month t – 2. The PBA approves prices not exceeding the price cap, which is equal to the highest existing price of exchangeable products, which implies that parallel imports are allowed to be as highly priced as locally sourced products [11, 12]. The price cap may prevent the seller of a product that is already the most expensive among its substitutes from increasing the price, if they want their product to remain within the benefit scheme. This means that a price cut retrospectively found to be too large cannot always be reversed. Before July 2009, producers and parallel importers were not allowed to offer their products below the prices approved by the PBA prices. That is, they were not allowed to give discounts to pharmacies.

Methods

Data

This study was based on two panel datasets obtained by merging datasets of pharmaceutical sales compiled by IMS Sweden (now part of IQVIA) with datasets containing detailed information of each pharmaceutical product, which were provided by the Västerbotten county council.5 An observation in the datasets represents a product with a certain active ingredient, strength, administrative form, and package size, supplied by a certain firm and sold in a certain month. The datasets cover all prescription drugs sold in Sweden during the periods of October 2002–October 2007 and January 2011–December 2017. Data from November 2007–June 2009 were not used because prices during this period could have been affected by anticipation of the possibilities of giving discounts. Because price increases are not always allowed within the benefit scheme, it would have been rational for firms to stop reducing list prices in response to competition upon discovering the possibility of giving discounts to pharmacies in the future. In this manner, they would have had higher list prices and therefore greater possibilities to give discounts upon legalization of the practice, compared to if they continued reacting to competition with lower list prices until the day discount was legalized.6 We also excluded data from July 2009–December 2010, as the business models related to discounts might still have been under rapid development under this period. Lacking information on patent expiration, we defined pharmaceuticals as off-patent starting from the first time any generics with the same active ingredient (i.e., the same 7-digit ATC code) were sold in Sweden, and pharmaceuticals are included in the analyses until the month they are designated to be off-patent. After excluding off-patent pharmaceuticals, the first and second datasets respectively contained 132,008 and 101,489 observations of locally sourced product and 31,999 and 70,540 observations of parallel-imported products. That is, for an average year, the first and second datasets respectively contained 25,969 and 14,498 observations of locally sourced products and 6295 and 10,077 observations of parallel-imported products.

Estimation of price effects and descriptive statistics

For several reasons, prices are not expected to adjust instantaneously to new long-term equilibriums when market conditions change. One reason is possible price coordination between therapeutic alternatives, which can cause companies to limit price changes to reduce the risk of triggering price wars [14]. Another reason is the dynamic price cap on drugs in Sweden, which means that a drug whose price is raised to a figure higher than that of the most expensive substitute can be excluded from the pharmaceutical benefit scheme. A company that is uncertain about what the new optimal price is after it has received competition may, because of this price cap, find it wise to lower the price gradually, rather than to lower it more directly and then risk not being able to adjust the price if it is found that the price cut was unnecessarily large. For these reasons, we estimated price effects with dynamic models.

The preferred specification, which was estimated with two-stage least squares using the STATA package xtivreg2, is written as:in which indices i, s, and t represent product, substance, and time in months, respectively. Variable definitions are presented in Table 1 and in the text below. The dependent variable lnP is the natural logarithm of the listed purchase price for all pharmacies for the on-patent locally sourced product i in month t. The first lag of this variable, lnP, was included as an explanatory variable to make the model dynamic.

Table 1

Variable definitions

lnPit	Natural logarithm of the listed purchase price for product i in month t.
D _ PiSubstancest	Equals one if one or more parallel imported product with the same active substance as product i were sold in Sweden in month t.
D _ Piit	Equals one if one or more parallel imported product exchangeable with product i was sold in month t.
N_PiSubstancest	The number of parallel traders that sold products with the same substance as product i in month t.
lnN_PiSubstancest	Natural logarithm of N_PiSubstancest when N_PiSubstancest >  0, but equal to zero when N_PiSubstancest = 0.
N_ Piit	The number of parallel traders that in month t sold products exchangeable with product i.
lnN _ Piit	Natural logarithm of N_Piit when N_Piit > 0, but equal to zero when N_Piit = 0.
D _ Thst	Takes the value of one if one or more other firm in month t sold a locally sourced product with the same five-digit ATC code as product i.
D _ ThGenst	Takes the value of one if, in Sweden, there existed a generic version of a substance that was sold in month t with the same five-digit ATC code as product i.
N _ Thit	The number of pharmaceutical substances with the same five-digit ATC code and with locally sourced drugs sold by firms other than the seller of product i in month t.
lnN _ Thit	Natural logarithm of N _ Thit when N _ Thit >  0, but equal to zero when N _ Thit = 0.
N _ ThGenst	Number of substances with generic versions and the same five-digit ATC code as product i for which generic as product i that was sold in month t.
lnN _ ThGenst	Natural logarithm of N _ ThGenst when N _ ThGenst >  0, but equal to zero when N _ ThGenst = 0.
Qst	The number of defined daily doses sold of products with substance s in month t.
lnQst	Natural logarithm of Qst.

The variable D_PiSubstance is an indicator that takes the value of 1 if one or more parallel-imported products with the same substance as product i were sold in Sweden in month t. D_PiE is also an indicator, but it only takes the value of 1 if at least one parallel imported product exchangeable with product i was sold in Sweden in month t. In accordance with the substitution rules, an exchangeable product was defined as a drug with the same active substance, form of administration, strength, and nearly identical package size.

The variable lnN_PiSubstance was defined as the natural logarithm of the number of parallel traders selling products with the same substance when N_PiSubstance is strictly positive, and takes the value of 0 otherwise. The variable lnN_PiE has the corresponding definition for the number of parallel traders selling exchangeable products.7 As the natural logarithm of one is zero, lnN_PiSubstance and lnN_PiE do not change values when the numbers of parallel traders selling products with the same substance and the number of traders selling exchangeable products, respectively, change from zero to one. Therefore, the coefficients for D _ PiSubstance and D_PiE capture the effects of the first parallel trader within the same substance and exchange groups, respectively, whereas the coefficient for lnN _ PiSubstance and lnN_PiE capture the effects of variations in strictly positive numbers of parallel traders.

The variables D _ Th − lnN _ ThGen were included to control for competition from firms that sold therapeutic alternatives, that is, products with other pharmaceutical substances that are intended for the same or similar medical diagnoses. D _ Th takes the value of 1 if at least one other firm sold a locally sourced product with the same five-digit ATC code in month t. If there was a generic version of at least one of these substances, also D _ ThGen takes the value of 1. The variable N _ Th (not included in the specification) was defined as the number of pharmaceutical substances with the same five-digit ATC code and with locally sourced drugs sold by firms other than the seller of product i during month t. lnN _ Th is the natural logarithm of N _ Th for strictly positive values of this variable and is otherwise 0. Lastly, lnN _ ThGen was defined as the natural logarithm of the number of therapeutic alternatives for which generic versions exist when this variable is strictly positive, and takes the value of 0 otherwise.

The eight competition variables were all instrumented with their first lags and with lnQ, which is the natural logarithm of the quantity of substance s sold in month t − 3.8 Producers have good information about the values of these instruments when, at the end of t − 2, they set their prices for month t. For the first eight instruments, the reason for this is that the prices of all products that can be sold within the benefit scheme in month t − 1 are announced in the first half of month t − 2. Hence, producers can observe how many potential competitors they will have in month t − 1 and can, based on this, predict the competition they will face in month t. Regarding lnQ, IMS/IQVIA had delivered sales data for month t − 3 to its customers when prices from month t were set. The validity of the instruments is discussed and analyzed in the Appendix. Lastly, month and product fixed effects (η and μ) were included in the specification, and the error terms were allowed to be correlated within substances.

To study if the functional form of the preferred specification was too restrictive, we also estimated a specification where D _ PiSubstance, D _ PiE, lnN _ PiSubstance, and lnN _ PiE were replaced by ten indicator variables for number of parallel trades within the substance and exchange group, respectively. In this specification, the lags of the ten indicator variables were used as instruments instead of D _ PiSubstance, D _ PiE, lnN _ PiSubstance, and lnN _ PiE; otherwise the specifications were identical to the preferred specification.

Descriptive statistics are presented in Table 2.

Table 2

Descriptive statistics

	Oct. 2002–Oct. 2007		Jan. 2011–Dec. 2017
	Mean	SD	Mean	SD	Min	Max
Pit	1462.75	4594.54	3762.37	12,019.43	6.31	290,670.50
D _ PiSubstancest	0.24	0.43	0.46	0.50	0	1
D _ PiEit	0.11	0.32	0.25	0.43	0	1
N _ PiSubstancest	0.62	1.36	1.52	2.11	0	11
N _ PiEit	0.24	0.82	0.61	1.34	0	9
D _ Thst	0.82	0.39	0.84	0.37	0	1
D _ ThGenst	0.53	0.50	0.54	0.50	0	1
N _ Thit	2.93	2.50	3.23	3.86	0	28
N _ ThGenst	0.89	1.15	1.11	1.41	0	8
Qst (in millions)	12.91	66.71	9.34	44.40	0.00	85,300.00

Note: The number of observations is 132,008 for the first dataset and 101,489 for the second dataset. See Table 1 for variable definitions

Regression results

The estimation results for the preferred specification are presented in Table 3, while model checks, results from ordinary least square regressions, and robustness analyses are presented in the Appendix. The results for the lag of the dependent variable (lnP) show that prices reacted slowly to changes in competition. Taking one minus the coefficient for lnP and multiplying by 100 reveals that only 4 and 8% of the long-term effects were realized immediately in the two sample, respectively.

Table 3

Estimation results for lnP

	Discounts forbidden	Discounts allowed
	Oct. 2002–Oct. 2007	Jan. 2011–Dec. 2017
lnPi, t − 1	0.9568***	0.9171***
	(0.0055)	(0.0174)
D _ PiSubstancest	−0.0017***	−0.0012
	(0.0006)	(0.0009)
D _ PiEit	−0.0012*	− 0.0003
	(0.0007)	(0.0005)
lnN _ PiSubstancest	0.0001	−0.0002
	(0.0004)	(0.0008)
lnN _ PiEit	−0.0010*	0.0009
	(0.0006)	(0.0008)
D _ Thst	−0.0001	0.0005
	(0.0008)	(0.0019)
D _ ThGenst	−0.0007	0.0002
	(0.0006)	(0.0012)
lnN _ Thit	−0.0012	−0.0026**
	(0.0012)	(0.0012)
lnN _ ThGenst	0.0010	−0.0011
	(0.0007)	(0.0015)
dlnPi*/dD_PiSubstancest*	−0.0601***	−0.0128
	(0.0151)	(0.0116)
Observations	119,945	90,228
R2	0.9183	0.8797
K-P rk LM	72.9704	65.2018
K-P rk LM, p-value	0.0000	0.0000
Hansen J, p-value	0.1293	0.1792

Note: See Table 1 for variable definitions. The specifications include product-specific fixed effects and 58 and 81 indicator variables for months, respectively. In the first-stage regressions, data from Oct. 2002–Oct. 2007 and Jan. 2011–Dec. 2017 were used. K-P rk LM refers to the Kleibergen-Paap rk LM statistic, which indicates the strength of the instruments. The null hypothesis in the K-P test is that the model is under-identified. The null hypothesis for the Hansen J test is that the instruments are valid, i.e., uncorrelated with the error term. Standard errors, robust to correlations within substances, are given in parentheses. ***, **, and * indicate that the coefficient is significantly different from zero on the 1, 5 and 10% significance levels, respectively. The estimation results for the indicator variables for months and for the first-stage regressions are available on request. In short, the first-stage regressions show that their own lag is the strongest instrument for each of the eight competition variables, with t-values ranging from 12 to 143 and point estimates from 0.67 to 0.93, and the R2 for the first stage regressions range from 0.49 to 0.89

The coefficients for the eight competition variables show their short-term effects, whereas dividing them by one minus the coefficient for lnP yields their long-term effects. To obtain the exact effect in percentage terms, the formula 100 ∗ [exp(B) − 1] should be applied, in which B is the coefficient estimate, or long-term estimate, of interest.

For the first study period, the estimates for D _ PiSubstance and lnP show that the first parallel trader selling products with the same active substance, but which were not exchangeable with product i, reduced the price of product i by 0.17% in the short-term and 3.9% [≈0.17/(1–0.9568)] in the long term. If the parallel trader instead sold an exchangeable product, so that D _ PiE also equaled one, the price fell by an additional 2.7% in the long term. Additional parallel importers only reduced the price if they sold exchangeable products, but in this case the effect was small as well; if the sellers of exchangeable products increased from one to three, the price was reduced by 2.2% in the long term.

The differential shows the weighted average long-term effect of facing competition from at least one parallel importer selling the same substance.9 Applying the formula 100 ∗ [exp(−0.0601) − 1], the effect equaled a price reduction by 5.83% for the first study period. In comparison, the raw (unweighted) average price reduction equaled 5.47%. The results are similar to those reported for tablets and capsules by Granlund [9]. For example, the long-term effect of a first parallel trader that also sold exchangeable products was estimated to be − 7.0% by Granlund [9], whereas here it was − 6.5%.

Column 3 of Table 3 shows that competition from parallel imports had no significant effect on the list prices of locally sourced products in the study period when discounts were allowed. Also, the weighted average effect of facing competition from at least one parallel importer selling the same substance was at the 5% level significantly smaller in the second study period compared with the first study period.

Figure 1 shows that similar results were obtained when instead using a more flexible specification with indicator variables for the numbers of parallel importers. However, for the period when discounts were forbidden, the confidence intervals were larger when indicator variables were used, which resulted in some estimates not being statistically significantly different from zero. For the period when discounts were allowed, the most notable difference was that the indicator variable for five parallel traders selling products with the same active substance was significantly different from zero, while no significant estimate was obtained using the preferred specification.

Fig. 1

Estimated long-term price effects in percentages of the number of parallel traders selling products with the same active substance and exchangeable products, respectively; comparison of logarithmic-form and flexible-form estimates. The effects in the left panels are plotted holding N _ PiE at zero, while the effects in the right panels are plotted holding N _ PiSubstance equal to N _ PiE, see Table 1 for variable definitions. The smooth lines are the long-term effects predicted from the preferred specification of D _ PiSubstance and lnN _ PiSubstance (left panels) and of D _ PiSubstance, D _ PiE, lnN _ PiSubstance, and lnN _ PiE (right panels). The gray area shows the associated 95% confidence intervals. Dummy point est. shows the long-term effects of indicator variables for the numbers of N _ PiSubstance (left panels) and for the numbers of N _ PiSubstance and N _ PiE (right panels), and Dummy CI, upper and Dummy CI, lower show the lower and upper bounds of the associated 95% confidence intervals. These estimates come from an IV specification including indicator variables for the numbers of parallel importers. However, groups with few observations were grouped together to avoid indicators that take the value of one for less than 1% of the observations. The estimates for these merged groups are plotted at the average value of N _ PiSubstance and N _ PiE in each merged group, respectively

Regarding the results for therapeutic competition, there were no statistically significant price effects in the first study period. For the second study period, the estimates for D _ Th and D _ ThGen imply that there was no significant price effect from a first therapeutic competitor, but the estimates for lnN _ Th and lnN _ ThGen indicate that prices fell with additional competitors.

Total savings of parallel imports when discounts were forbidden

The pharmaceutical costs in the absence of parallel imports were calculated by multiplying the sold quantity of all products (both locally sourced products and parallel imports) by the price that locally sourced products would have had in the absence of parallel imports.10 The savings are the difference between this cost and the actual cost. The savings were calculated using data from January 2003–October 2007 (i.e., for the period used in the second stage of the IV regressions) and divided into one direct and two indirect parts; one for locally sourced products and one for parallel-imported products.

The direct savings consists of the sum over all parallel-imported products of the number of packages sold multiplied by the price difference between these and their locally sourced counterparts. Parallel imports were on average 9% cheaper than locally sourced products, yielding an average annual direct saving of 231 million SEK (24 million Euros) in 2017 prices.11 After discounts were allowed, the gaps in list prices between parallel imports and locally sourced drugs were on average only 0.8%.12

The indirect savings for locally sourced products were calculated as the total sales value of locally sourced products for which D _ PiSubstance = 1, multiplied by 0.0619, which shows in decimal form the estimate of how much more expensive the products would have been if they had not faced competition from parallel imports.13 These savings were estimated to average 260 million SEK (27 million EUR) per year. The indirect savings for parallel-imported products were calculated correspondingly, except that we used the sales values that would have existed if these products had been sold at the same price as the locally sourced products. This yielded an estimated average indirect savings for the parallel imports of 161 million SEK (17 million EUR) per year.

The savings are illustrated in Fig. 2 and summarized in Table 4. The rectangle of Fig. 2 illustrates how large the annual expenditure on patent prescription pharmaceutical was estimated to have been without competition from parallel imports. For locally sourced products that did not face competition from parallel imports, no savings occurred. The savings for the other two categories are illustrated in the upper right corner of the figure. All in all, the estimated annual savings generated by parallel imports of pharmaceuticals before discounts were allowed totaled 652 million SEK (≈ 260 + 161 + 231) (68 million EUR). This amounts to 4% of the 16.619 billion SEK that on-patent prescription pharmaceuticals were predicted to have cost without parallel imports.

Fig. 2

Illustration of average yearly savings for January 2003–October 2007. LS refers to locally sourced products; PI refers to parallel-imported products. The amounts are measured in million SEK of pharmacies’ purchase prices and are expressed in year 2017 prices. In 2017, the average exchange rate was 9.64 SEK for one EUR

Table 4

Predicted savings with CIs, in millions SEK

	Point estimate	95% CI estimation uncertainty	95% CI PSA
Direct savings	231		113–375
Indirect LS	260	130–394	134–424
Indirect PI	161	81–245	81–271
Indirect LS + PI	421	211–638	223–679
Direct + indirect PI	393	312–476	238–577
Total savings	652	442–869	391–968

Note: The asymmetry in the CIs in Column 3 is explained by the concavity of C/(100 - C) (described in footnote 13), which is only partly offset by the convexity of C = 100*[exp(B)-1]. For the PSA CIs, the randomness of the Monte Carlo simulations is also a source of asymmetry. In 2017, the average exchange rate was 9.64 SEK for one EUR

In Table 4, Column 2 reports the point estimates, and Column 3 reports the 95% confidence intervals (95% CI) that reflect uncertainty in the estimated price effects of competition from parallel imports only. The last column reports the CIs from a probabilistic sensitivity analysis (PSA) that also accounted for variability in competition from parallel imports, market shares, and relative prices. The CIs are further described and discussed in the Appendix.

Discussion and conclusions

When discounts were forbidden, parallel imports had a market share of 16% and were on average 9% cheaper than locally sourced drugs, which directly reduced the cost for on-patent pharmaceuticals by 1.4%. Additionally, we estimated that parallel imports reduced the prices of products with the same substance by, on average, 6% in the long-term. Combining this with the share facing competition from parallel imports indicates that, in total, parallel imports reduced the cost of on-patent pharmaceuticals by 4%.

The estimated price-effects of competition from parallel imports are significantly lower than reported by previous studies [3-5] that used possible endogenous instruments, such as exchange rates. A main advantage of the dynamic model used in this study is that the lagged dependent variable controls for previous price shocks, which makes lagged values of the competition variables valid instruments for the current values. This provides enough strong instruments to estimate the price effects of competition from parallel imports on both the extensive and intensive margin and to do this using flexible specifications. A drawback of the dynamic model is that including a lag of the dependent variable in models with fixed effects can result in bias [16]. As described in the Appendix, this bias is expected to be very small for datasets with high numbers of time periods, such as those used in this study. For short study periods, researchers might have to address this problem by using methods such as an Arellano-Bond estimator [17], which can, in turn, make it difficult to find strong instruments for the competition variables.

For the period when discounts were forbidden, we found no statistically significant effects on list prices despite narrow confidence intervals. This strengthens the conjecture that the preference of locally sourced product sellers to compete with parallel imports by giving discounts to pharmacies is a main reason that previous studies [6-8] have not found significant effects on list prices when discounts were legal. Some advantages of using discounts are that they can be quickly reverted and do not affect the maximum prices producers are allowed to charge in countries that use external reference pricing [18].

No estimates exist regarding the discounts given by sellers of locally sourced products, but the discounts given by parallel traders have been estimated to be about 470 million SEK per year (49 million EUR).14 This is within the CI of savings in the pre-reform period caused by parallel imports having lower prices than their locally sourced counterparts would have had if their prices had not been lowered due to competition from parallel imports (point estimate, 393 million SEK). It is conceivable that allowing discounts had small effects on the total savings caused by parallel imports, but allowed those saving to go to pharmacies rather than to consumers and the pharmaceutical benefit scheme. In the case of Sweden, the government can easily redirect savings from pharmacies to consumers and the pharmaceutical benefit scheme by changing the formula that dictates how high the prices pharmacies charge should be in relation to the list prices they pay when not receiving discounts.

From the results of this study, one cannot draw any conclusion on whether it is preferable to allow or forbid sellers of on-patent pharmaceutical to give discounts to pharmacies, as the savings caused by parallel imports can be equally large under both conditions. However, the results clearly show that, from the perspective of payers in destination countries, it is beneficial to continue allowing parallel imports. Likewise, introducing rules that are hard to meet for parallel imports—for example that firms selling pharmaceuticals to pharmacies need to have large quantities in stock, as suggested by a Swedish government inquiry—will be costly. From a global welfare perspective, it is harder to draw policy conclusions regarding parallel trade because the savings of consumers and insurances come at the expense of reduced revenues for sellers of locally sourced drugs. To reduce the amount of parallel trade, these sellers might also increase prices and delay launch of new drugs in low-income countries, which could cause welfare losses [22]. Theoretical studies indicate that the total welfare effects of allowing parallel trade with pharmaceuticals are generally ambiguous and partly depend on differences in national pricing rules [23, 24], patients’ preferences [23, 25] and the vertical integration of trading companies [26].

Table 5

Robustness checks for Oct. 2002–Oct. 2007, estimation results for lnP

	IV 2	IV 1 s2	IV 1 s2 cl5	OLS s2	Static OLS s2
lnP i, t − 1	0.9568***	0.9568***	0.9568***	0.9568***
	(0.0055)	(0.0055)	(0.0057)	(0.0055)
D _ PiSubstancest	−0.0019***	− 0.0017***	− 0.0017***	− 0.0015***	−0.0106**
	(0.0007)	(0.0006)	(0.0006)	(0.0005)	(0.0045)
D _ PiEit	−0.0012	−0.0012*	− 0.0012	−0.0010*	0.0068
	(0.0008)	(0.0007)	(0.0007)	(0.0005)	(0.0049)
lnN _ PiSubstancest	0.0000	0.0001	0.0001	0.0001	0.0057
	(0.0004)	(0.0004)	(0.0004)	(0.0003)	(0.0057)
lnN _ PiEit	−0.0012*	−0.0011*	−0.0011*	− 0.0007	0.0298***
	(0.0007)	(0.0006)	(0.0006)	(0.0005)	(0.0110)
D _ Thst	−0.0002	−0.0001	− 0.0001	− 0.0002	0.0040
	(0.0010)	(0.0009)	(0.0009)	(0.0007)	(0.0099)
D _ ThGenst	−0.0012**	−0.0007	− 0.0007	− 0.0002	− 0.0127***
	(0.0006)	(0.0006)	(0.0006)	(0.0003)	(0.0045)
lnN _ Thit	−0.0003	−0.0012	− 0.0012	−0.0004	− 0.0049
	(0.0013)	(0.0012)	(0.0013)	(0.0010)	(0.0127)
lnN _ ThGenst	0.0010	0.0009	0.0009	0.0007	−0.0199**
	(0.0008)	(0.0007)	(0.0007)	(0.0006)	(0.0095)
dlnPi*/dD_PiSubstancest*	−0.0671***	−0.0603***	− 0.0603***	− 0.0496***	0.0067
	(0.0165)	(0.0151)	(0.0156)	(0.0129)	(0.0087)
Observations	119,058	119,058	119,058	119,058	119,058
R2	0.9181	0.9181	0.9181	0.9181	0.0201
K-P rk LM	69.1940	72.9573	41.6238
K-P rk LM, p-val.	0.0000	0.0000	0.0000
Hansen J, p-value	0.1496	0.1346	0.1439

Note: See Table 1 for variable definitions. The specifications include product-specific fixed effects and indicator variables for year × month combinations. In the first-stage regressions, data from Oct. 2002–Oct. 2007 and Jan. 2011–Dec. 2017 were used. K-P rk LM refers to the Kleibergen-Paap rk LM statistic, which indicates the strength of the instruments. The null hypothesis in the K-P test is that the model is under-identified. The null hypothesis for the Hansen J test is that the instruments are valid, i.e., uncorrelated with the error term. Standard errors, robust to correlations within substances, are given in parentheses. ***, **, and * indicate that the coefficient is statistically significant different from zero on the 1, 5 and 10% significance levels, respectively. The estimation results for the indicator variables for year × month combinations and for the first-stage regression are available on request

Table 6

Robustness checks for Jan. 2011–Dec. 2017, estimation results for lnP

	IV 2	IV 1 s2	IV 1 s2 cl5	OLS s2	Static OLS s2
lnP i, t − 1	0.9193***	0.9194***	0.9194***	0.9194***
	(0.0177)	(0.0176)	(0.0202)	(0.0176)
D _ PiSubstancest	−0.0011	− 0.0012	− 0.0012	−0.0007	− 0.0069
	(0.0010)	(0.0009)	(0.0009)	(0.0005)	(0.0070)
D _ PiEit	−0.0005	−0.0003	− 0.0003	−0.0006	0.0022
	(0.0006)	(0.0005)	(0.0005)	(0.0004)	(0.0046)
lnN _ PiSubstancest	−0.0003	−0.0001	− 0.0001	−0.0001	0.0018
	(0.0009)	(0.0008)	(0.0009)	(0.0006)	(0.0087)
lnN _ PiEit	0.0010	0.0009	0.0009	0.0009	0.0161***
	(0.0009)	(0.0008)	(0.0007)	(0.0007)	(0.0061)
D _ Thst	0.0021	0.0010	0.0010	−0.0000	−0.0235
	(0.0020)	(0.0017)	(0.0018)	(0.0011)	(0.0209)
D _ ThGenst	0.0004	0.0003	0.0003	0.0000	0.0115
	(0.0012)	(0.0012)	(0.0014)	(0.0010)	(0.0119)
lnN _ Thit	−0.0037**	−0.0025**	− 0.0025*	−0.0011	− 0.0183*
	(0.0015)	(0.0012)	(0.0014)	(0.0009)	(0.0101)
lnN _ ThGenst	−0.0010	−0.0012	− 0.0012	−0.0011	− 0.0091
	(0.0017)	(0.0015)	(0.0017)	(0.0013)	(0.0128)
dlnPi*/dD_PiSubstancest*	−0.0152	−0.0132	− 0.0132	−0.0085	0.0055
	(0.0122)	(0.0112)	(0.0128)	(0.0107)	(0.0104)
Observations	89,292	89,292	89,292	89,292	89,292
R2	0.8830	0.8831	0.8831	0.8831	0.0122
K-P rk LM	72.1080	65.6942	24.1794
K-P rk LM, p-val.	0.0000	0.0000	0.0000
Hansen J, p-value	0.1830	0.1639	0.1950

Note: See Table 1 for variable definitions. The specifications include product-specific fixed effects and indicator variables for year × month combinations. In the first-stage regressions, data from October 2002–October 2007 and Jan. 2011–Dec. 2017 were used. K-P rk LM refers to the Kleibergen-Paap rk LM statistic, which indicates the strength of the instruments. The null hypothesis in the K-P test is that the model is under-identified. The null hypothesis for the Hansen J test is that the instruments are valid, i.e., uncorrelated with the error term. Standard errors, robust to correlations within substances, are given in parentheses. ***, **, and * indicate that the coefficient is statistically significant different from zero on the 1, 5 and 10% significance levels, respectively. The estimation results for the indicator variables for year × month combinations and for the first-stage regression are available on request