Physicians' Prescribing Responses to a Restricted Formulary: The Impact of Medicaid Preferred Drug Lists in Illinois and Louisiana

January 15, 2005
Suchin Virabhak, PhD

Judith A. Shinogle, PhD

Volume 11, Issue 1 SP

Objective: To determine the extent to which the strictness of the criteria used to control utilization of drugs not on a Medicaid preferred drug list (PDL) changes physician prescribing habits for non-Medicaid patients.

Study Design: Quasi-experimental design based on different timing of states'enactment of PDLs for cardiovascular medications.

Methods: A regression model was developed to analyze the effect that PDL implementation had on off-PDL product market share across 3 patient types (Medicaid, third-party insurance, and cash paying). The models included data from 2 states (Illinois and Louisiana) with different PDL prior-authorization criteria. The data allowed examination of different physician responses according to the proportion of Medicaid patients treated by the physician. The analysis also followed prescribing patterns longitudinally to determine whether the PDL-induced prescribing behaviors changed over time.

Results: There was a decrease of 9 percentage points (67.7%) and 6.2 percentage points (40.5%) in the Medicaid prescription share of restricted cardiovascular drugs post-PDL for Illinois and Louisiana, respectively. In the third-party insurance market, prescription shares of off-PDL drugs decreased 0.9 percentage points (6.8%) in Illinois and 1.3 percentage points (8.6%) in Louisiana. For physicians with a high percentage of prescriptions paid for by Medicaid, the share loss for off-PDL drugs was estimated to be more than 37% for the non-Medicaid portion of the practice.

Conclusion: The effects of a Medicaid PDL on prescribing behavior extend beyond the Medicaid population. The health outcomes and economic consequences of these "spillover" effects are poorly understood and warrant further research.

(Am J Manag Care. 2005;11:SP14-



Medicaid prescription drug expenditures grew from $4.4 billion in 1990 to more than $20 billion in 2000, representing an average annual increase of 16.3%.1 Faced with lower revenues and thus tighter budgets, many state Medicaid programs have introduced preferred drug lists (PDLs) as a means of cost control. Preferred drug lists designate specific drugs for use by beneficiaries while requiring prior authorization from Medicaid to access drugs not included on the list (off-PDL drugs). When physicians comply with PDL restrictions, by definition they change their prescribing habits toward Medicaid patients. Moreover, to the extent that PDLs affect prescribing patterns, non-Medicaid patients are indirectly impacted. We focus on the latter effect in this paper. Should spillovers in prescribing behavior occur, it would be worthwhile for policy analysts to study the long-term effect of PDLs on health outcomes of both the Medicaid and non-Medicaid populations.

In this study, we examined 2 states that implemented Medicaid PDLs in 2002 along with 2 control states, and estimated changes in physicians' prescribing behavior in the Medicaid market as well as potential spillover effects in other markets. Our approach enabled us to examine policy-induced changes in physician prescribing behavior more precisely than in previous studies, with a particular focus on whether there are large spillover effects in the non-Medicaid market for physicians who treat a relatively high proportion of Medicaid patients.


In mid-April 2002, Illinois authorized the use of a Medicaid PDL; by mid-June 2002, calcium channel blockers became the first group of cardiovascular drugs to have a PDL. In early July 2002, it authorized PDLs for angiotensin-converting enzyme (ACE) inhibitors and angiotensin receptor blockers. Preferred drug lists for beta blockers and alpha blockers were introduced later, in November 2002 and February 2003, respectively. Centrally acting agents, diuretics, and spirolactone were completely unaffected by the PDL. In Illinois, the onus of verifying a patient's eligibility and obtaining prior authorization rests on an enrolled provider. Specifically, the "physician must document that the drug requested is necessary to prevent a life threatening situation and that items covered without prior authorization?are not effective." When a request is made for a brand-name drug for which a generic version exists, "medical information regarding the lack of effectiveness of the generic" must be provided.2

Louisiana implemented a Medicaid PDL for cardiovascular drugs in August 2002 that included ACE inhibitors, angiotensin receptor blockers, and calcium channel blockers. Preferred drug lists for beta blockers and alpha blockers were phased in later, in September and October, respectively. Diuretics, spirolactone, and centrally acting agents were not affected by the Medicaid PDL. Physicians are responsible for submitting prior-authorization requests but do not face the same degree of stringency as in Illinois (eg, including verification of eligibility or providing evidence of the suitability of a nonpreferred treatment).

Together, Illinois and Louisiana provide a fertile opportunity to analyze changes, if any, in physicians' prescribing behavior and resulting spillover effects after enactment of a Medicaid PDL. In addition, the PDLs in both states were implemented more than a year ago, which allows sufficient time for habits to change. A difference in the magnitude of change in prescribing behavior between the states may reflect differences in severity of the prior-authorization procedures.


Previous research on physician behavior has found that treatment patterns are most distinct at the practice level (not the patient level), where treatment intensity depends on the insurance status of patients treated by the same physician as much as the insurance status of an individual.3 After Maine's Medicaid program implemented a restrictive formulary that included only 1 drug (pantoprazole) in the proton pump inhibitor class, Wang et al used data from October 1999 to September 2001 and a difference-in-difference-in-differences model at the physician level to examine the impact of this change on the market share of pantoprazole. Their evidence suggests that the PDL generated spillover effects in the order of 1.4% to 1.8% for non- Medicaid patients for each 10% increase in the physician's Medicaid share, depending on the payment type.3

Our study adheres to the methodology of Wang et al. We used state controls and focused on all the drugs within a relatively broad therapeutic class (ie, cardiovascular drugs) rather than just 1 drug.4 We expanded on their analysis by tracking the introduction of PDLs in 2 states instead of 1 state, which allowed us to compare the effects of PDL strictness on prescribing behavior. The analysis also separated the post-PDL period into subperiods to shed light on whether prescribing behavior undergoes short-term "undershooting"or "overshooting," as suggested by Grabowski.5 Undershooting implies that the change in behavior grows over time, whereas overshooting implies that physicians initially change their behavior but tend to revert back to old habits over time despite the PDL. The time element also is important because PDLs often permit a fixed number of refills to be exempted from prior authorization. In Louisiana, for instance, up to 5 refills are exempted. Therefore, allowing sufficient time to elapse will reflect changes in prescriptions when refill limits are reached.

Our hypothesis–consistent with the findings of Wang et al for a specific drug–is that physicians serving a relatively large share of Medicaid patients will be more affected by Medicaid PDLs, and that spillover effects on their non-Medicaid patients will be larger. We expected spillover effects to be greater among patients with third-party coverage, because cash paying-patients are price sensitive and are more likely to request a low-cost drug. Furthermore, physicians may view cash-paying patients to be poorer than covered patients, a perception that may encourage them to prescribe lower-cost drugs. We also hypothesized that the state with the stricter prior-authorization regime would have a greater impact on physician behavior in the Medicaid market.


We used a quasi-experimental design based on different timing of states'enactment of PDLs compared with similar states without PDLs to examine the aggregate effect of physicians' responses to a PDL for cardiovascular medications (see the Background section for a list of the affected drugs). The analysis was restricted to those physicians who had experience writing prescriptions in the therapeutic category of interest. whom we termed "high-script physicians. "High-script physicians included those physicians in the top third of all prescribers of cardiovascular medications. High-script physicians are likely to develop consistent prescribing habits, whereas low-script physicians, due to lack of sufficient exposure, may have more variable prescribing habits. In contrast to Wang et al, we conducted panel analysis at the aggregate market level rather than physician level. Moulton showed that if the disturbances are correlated within the groupings that are used to merge aggregate with micro data, even small levels of correlation can cause standard errors from ordinary least squares to be seriously biased downward, which can result in spurious findings of statistical significance for the aggregate variable of interest.6




We measured prescribing habits of physicians before and after PDL implementation by observing the trend in Medicaid prescription share of off-PDL drugs. In order to capture spillovers, we observed the trend in cash-paying and third-party markets. For each market (cash, third party, and Medicaid), state , and time , our variable of interest was:



where is the share of off-PDL drugs and is the number of new prescriptions. Drugs included on the PDL are type ON, whereas those not specifically included or excluded on the PDL are termed NIL.


Omitted variables and unobserved heterogeneity may have confounded our results. For example, an adverse medical report that coincidentally might have steered physicians away from off-PDL drugs could be mistaken for a market response to PDL. To control for this possibility, we used a difference-in-differences approach by including control states. We measured the impact of the PDL using an indicator variable for whether the PDL was in place in a given state at time Thus, the model we estimated is:







_is defined similarly. is a vector of state-specific binary controls including the mean monthly share of off-PDL prescriptions during the pre-PDL period in market ; dtcaptures time-fixed effects, in months; emjtis a disturbance term. Each observation is weighted by the average number of monthly pre-PDL prescriptions in that state. The direct impact of the PDL was measured in the Medicaid market, where ßis associated with the legislation-induced change in prescription shares of off-PDL drugs in Illinois and ßis the corresponding impact in Louisiana. If our spillover hypothesis is true, there should be a similar, albeit weaker, outcome in the cash-paying and third-party payer markets.





We extended equation 2 to capture differences arising from Medicaid share of practice. This extension was introduced by simply dividing the data by a covariate that captures the Medicaid share of practice (henceforth ). was computed at the physician level as the ratio of Medicaid to total prescriptions in the baseline (ie, pre-PDL) period. It was important to use baseline measures to eliminate the possibility that this variable was endogenous to PDL implementation. We divided high-script physicians into 4 categories: NONE, LOW, MED, and HIGH, with respective pre-PDL Medicaid practice shares of 0%, >0% to <5%, 5%-50%, and >50%. (Our results were not highly sensitive to various categorizations or a linear specification.) We then aggregated our data by . The model became:






Illinois, Louisiana




where _is a dummy variable for Medicaid share of practice and is an indicator for state of interest ?{}. _, = NONE was excluded so that coefficients of the dummies were relative to this category. The interaction term distinguished the impact of the PDL on each practice type. We hypothesize for ßs: |ßHIGH| > | ßLOW|.

Finally, we divided the data into subperiods with short-term effects in the first 6 months post-PDL, and long-term effects in the first year post-implementation. Physicians are informed about a new formulary about a month before implementation and may respond immediately. Hence, a 2-month period defines implementation. We compared the difference in prescription shares between a state and its control in each period using time dummies to test for differences in short-term and long-term changes in prescribing behavior. Due to the lack of data, we were unable to control for "bandwagon effects" (where private insurers mimic Medicaid formulary changes) that may have confounded long-term spillover effects. Our models measure prescription shares, which by definition are limited to values between 0 and 1. Ideally, we should use a model that incorporates the distributional aspects, but to ease interpretation of results (especially interactions), we applied ordinary least squares estimation.


We obtained prescriber-level and payer/prescriber-level data from NDCHealth (Atlanta, Ga) for the period from January 2002 to September 2003 in Illinois and New York, and for January 2002 to May 2003 in Louisiana and Mississippi. Physicians with at least 1 cardiovascular prescription during the pre-PDL period were included. NDCHealth samples more than 70% of the total number of dispensed prescriptions from retail pharmacies, 98% of which are matched back to their respective prescribers. Our data included 23 111 physicians from Illinois, 27 844 physicians from New York, 8996 physicians from Louisiana, and 4525 physicians from Mississippi. For each drug product, monthly new prescriptions by market (cash, Medicaid, or third party) were recorded. Cash payers in the NDCHealth data include any script paid in full by a patient at a pharmacy; not all of these are true cash payers, however, as some scripts are reimbursed later. A study of statins showed that about 35% of cash payers have prescription drug coverage (IMS Custom Analysis, Fairfield, Conn). A new prescription is not equivalent to a new patient, but it is a newly originated prescription (as opposed to a refill). Since refills of existing prescriptions are not necessarily affected by a PDL, new prescriptions would most accurately reflect changes in prescribing habits.


Table 1 presents descriptive statistics for selected population and economic indicators in Illinois and Louisiana, and compares these with their respective controls, New York and Mississippi. The states and their controls were selected so that demographics such as race and per capita income were closely matched. At baseline, Medicaid's share of cardiovascular prescriptions was approximately 10% higher in the control states. However, we calculated the shares of off-PDL new prescriptions by market type (which were comparable across states); thus, the differing baselines should not bias the analysis. The resulting change in off-PDL drug market share in the pre-PDL and post-PDL periods across markets is displayed in Table 2. Across all 3 markets, the average prescription shares of off-PDL drugs were fairly close between states in the pre-PDL period. Post-PDL, shares fell steeply in the PDL states in the Medicaid market, despite corresponding small increases in non-PDL states. Spillover effects appeared weaker in the third-party market and were not immediately discernible in the cash market. Figure 1 illustrates the trend in off-PDL prescription shares across time, and reinforces the preceding findings. Moreover, Figure 1 hints that Medicaid shares of new cardiovascular prescriptions in Illinois may have increased slightly about a year after implementation of the PDL (July 2003).


Figure 2A and Figure 2B show what happened when practice share of Medicaid was controlled for () in Illinois. In this state, spillovers into the third-party and cash-payer markets were especially pronounced and enduring for the high-script physicians, and short-term undershooting was evident. Prescription shares returned to their pre-PDL values for the remaining groups of physicians in the third-party market, which indicated short-term overshooting; evidence of spillovers was not apparent for both no-script and low-script physicians in the cash market. (We calculated the corresponding graphs for Louisiana; they mirror those of Illinois, with a less precipitous fall in prescription shares. Figures are available upon request.)

Results from our regression analysis are presented in Table 3. Column 1 of Table 3 presents our results for the Medicaid market: the legislation-induced change in off- PDL prescription shares was -9.0 percentage points in Illinois and -6.2 percentage points in Louisiana. Given that the average off-PDL Medicaid share in the pre-PDL period was 13.5% and 15.3%, respectively, this amounts to a 67.7% decrease in the average pre-PDL Medicaid prescription share in Illinois and a 40.5% decrease in Louisiana. The difference between the 2 states is 2.8 percentage points and is statistically significant (7%), which suggests the larger impact of a stricter prior-authorization procedure. Column 2 displays the results of estimating the effects of physicians'Medicaid practice share. The 3.1% difference between Illinois and Louisiana in the legislation-induced impact of the PDLs is close to our previous estimate.

The interaction term reveals that in both Illinois and Louisiana, the fall in off-PDL prescription share was larger for physicians who saw a greater percentage of Medicaid patients. The LOW (practices with >0% but <5% Medicaid patients) and MED (practices with 5%-50% Medicaid patients) types were not statistically different from NONE (practices with 0% Medicaid patients). However, with the HIGH type (practices with >50% Medicaid patients) there was a significant 4.7 percentage point additional decrease in prescription share. This amounts to a marginal fall in off-PDL prescription share of about 10% in Illinois and 13.1% in Louisiana, which translates to a fall in the average pre-PDL prescription share of 80% in Illinois and 60% in Louisiana for Medicaid prescribers in the HIGH group. We found no evidence of any significant difference between Illinois and Louisiana for each group of physicians (results not shown).

Columns 3 and 4 tabulate the spillover results for the third-party and cash markets, respectively. As expected, the direct impact of the PDLs exceeded their spillover effects. The legislation-induced decrease in third party off-PDL prescription share was small but statistically significant: -0.9 percentage points for Illinois and -1.3 percentage points for Louisiana, which amounts to a 6.8% decrease in the average pre-PDL third-party prescription share for Illinois and a 8.6% decrease for Louisiana. The difference between both states was not significant at 10%. On the other hand, the legislation-induced decrease in off-PDL prescription share in the cash market was -0.4 percentage points in Illinois; there was no statistical evidence of an impact in Louisiana, nor a difference between both states.

To the extent that a given type of cardiovascular drug class is designed to treat certain indications, switches in drug regimens will occur within a drug class. Furthermore, certain drugs may have potential unfavorable effects. In an extension, prescription shares of off-PDL drugs were computed within each drug class. After implementation of the PDL, the Medicaid prescription share in Illinois fell by 10.1%, third-party share prescription by 0.9%, and cash prescription share by 0.6%; the comparable figures for Louisiana are 6.3%, 0.8%, and 0%. These results are qualitatively the same as those found previously at the market level for cardiovascular drugs; moreover, the estimates are close to those previously obtained. This finding can be interpreted as empirical support for the supposition that switches in the cardiovascular market after a PDL are essentially between drugs within the same class. Complete results are available from authors upon request.

Given that a majority of physicians in our sample had Medicaid practice shares of more than 5% (MED and HIGH types), we ran a model similar to the practice-share model using this restricted sample. Our results for the third-party payer market are reported in column 5 of Table 3. The legislation-induced reduction in prescription share was 2.3% in both Illinois and Louisiana; moreover, HIGH practices registered an additional 3.7 percentage point reduction in prescription share relative to MED practices, and the difference is statistically significant. Hence, the estimated third-party spillover impact in the HIGH practices was -6 percentage points, or equivalently, a 37.5% decrease in the average pre-PDL share in both states. In the cash market (Table 3, column 6) legislation-induced spillovers for HIGH relative to MED practices were -2.5 percentage points in Illinois and -0.3 percentage points in Louisiana (not significantly different from zero).

We did not find support for Grabowski's hypothesis of major long-term adjustments in prescription shares in the Medicaid market. In the third-party payer market, there was a significant fall in the short term, but no evidence of a difference in prescription shares between the short and long term. In the cash market, there was a significant but small 0.4% short-term increase in prescription shares; in the long term, prescription shares fell by 0.4 percentage points. Detailed results are available from the authors upon request.


Sensitivity analysis was performed on the models to control for clustering or prescribing around a physician. A model at the physician level that accounts for physician fixed effects was analyzed. The qualitative results were unchanged; as conjectured, standard errors were considerably smaller, which enhance the estimated values. Results are available from the authors upon request.


We found significant evidence of relatively large direct and spillover effects of Medicaid PDLs across Medicaid, third-party payer, and cash markets. Our estimates imply that the legislation-induced decline in the average pre-PDL Medicaid prescription share was 67.7% in the stricter program enacted in Illinois and 40.5% in the less binding program in Louisiana. This suggests that a more restrictive prior-authorization procedure leads to a greater reduction in off-PDL drug prescription shares in Medicaid, but it should be noted that even in the more lenient state the effect on physician behavior is significant. Moreover, physicians who see a greater proportion of Medicaid patients reacted more strongly to a Medicaid PDL. For physicians whose practices were more than 50% Medicaid, average third-party prescription shares of off-PDL products fell by 37.5%. We also found strong evidence of undershooting in the Medicaid market, as off-PDL drug prescription shares fell after 6 months of implementation and even further after a year. Future research should examine the effect of Medicaid PDLs on the benefit and formulary changes of private insurers to determine whether bandwagon jumping occurs.

In short, this research demonstrates that Medicaid drug cost-containment strategies such as PDLs appear to have far-reaching effects not only on the Medicaid population but also on privately insured patients. Such effects can be reduced by a less stringent prior-authorization procedure. Our present study does not examine whether the legislation-induced changes in physician behavior lead to worse health outcomes, an important issue that warrants further research.