1. SID, SAV, and Efficient Use Chapter 15 © Allen C. Goodman, 2013

2. Supplier Induced Demand Does more providers  more treatment? S1S1 S2S2 D1D1 D2D2 D'2D'2 Price of Services Quantity of Services P1P1 Q1Q1 Suppose S  from S 1 to S 2. You would expect changes to P 3, Q 3. Q3Q3 P3P3 If physicians can induce demand, however, to D 2, or D' 2, they can avoid losses. What kind of evidence do we need.

3. Big Econometric Identification Problem Start with model without SID (From earlier editions – not in FGS/5 – 7) Q D = a 0 + a 1 P + a 2 Y + u 1 (10.1) Q S = b 0 + b 1 P + b 2 X + b 3 MD + u 2 (10.2) where MD is number of MDs. In Eq’m: a 0 + a 1 P + a 2 Y + u 1 = Q D = Q S = b 0 + b 1 P + b 2 X + b 3 MD + u 2 or P = (b 0 -a 0 )/(a 1 -b 1 ) - a 2 Y /(a 1 -b 1 ) +b 2 X/ (a 1 -b 1 ) + b 3 MD/ (a 1 -b 1 ) Substituting into either (10.1) or (10.2) Q = c 0 + c 2 X + c 3 Y + c 4 MD + v(10.3) This looks like SID, except there was no SID in the model.

4. Why not induce all the time? Perhaps, inducement is a “bad.” Gives us unusual indifference curves. We see U 1 > U 2. Inducement, I Net Income U1U1 U 2 < U 1 I2I2 I1I1 p = mQ 0 + mI Suppose m falls (due to increased competition). It could give us less inducement … or more inducement. mQ 0 m Q0m Q0 I3I3 U' 2 < U 1 m = profit rate Can not have BOTH U 2 and U' 2

5. Let’s decompose a decrease in m into income and substitution effects Suppose m falls Inducement, I Net Income U1U1 U 2 < U 1 I2I2 I1I1 p = mQ 0 + mI Income effect – Drop in m implies more inducement Substitution effect – Drop in m makes inducement less effective. We do less inducement. Income and Substitution Effects Income Effect Substitution Effect m = profit rate

6. Suggests that for there to be major increase in inducement, income effects must dominate substitution effects. Research findings, at this point are varied. Inducement, I Net Income U1U1 U 2 < U 1 I2I2 I1I1 p = mQ 0 + mI Income and Substitution Effects Income Effect Substitution Effect m = profit rate

7. Important Issue If you believe in SID, then demand-side policies have little impact because providers can always induce more demand. Some people argue that of course providers induce demand – if so, so do mechanics.

8. SAV – Small Area Variation Is the Right Amount of Treatment Used? Usage of technologies may vary. Why? –Provider may not have complete knowledge of patient’s condition. –May not have complete knowledge of appropriateness of procedure. –Provider may have preferences among types of treatment. –Patient may have preferences among types of treatment.

9. Wennberg Practice style -- Why do practices vary so much? Medical Care Health Status S1S1 S2S2 S* Medical Care Health StatusS1S1 S2S2 S* Mgl. Cost Phys. 2 is shown as believing that additional units of medical care are more effective than does Phys. 1 Rate w/in a market depends on distributions of type 1 and type 2 Phys. M1M1 M2M2 Mgl. Product May vary within same office!

10. An Example Cesarean sections. Reference: Dartmouth Atlas for Michigan, Pp 8-9.Dartmouth Atlas for Michigan What does it mean?

11. McAllen v. El Paso Both cities on Rio Grande River. Both with large Hispanic percentages. McAllen – 89% El Paso – 82% Source: Franzini et al.Franzini et al.

12. Gawande - 2009 What’s going on? Gawande A. The cost conundrum.New Yorker [serial on the Internet]. 2009 Jun 1 [cited 2010 Nov 3].Available from: http://www.newyorker.com/reporting/2009/06/01/090601fa_fact_gawande

13. Franzini, Mikhail, Skinner Look at medical use and expense data for patients privately insured by Blue Cross and Blue Shield of Texas. In contrast to the Medicare, use of and spending per capita for medical services by privately insured populations in McAllen and El Paso was much less divergent, with some exceptions. Although spending per Medicare member per year was 86% higher in McAllen than in El Paso, total spending per member per year in McAllen was 7% lower than in El Paso for the population insured by BCBS of Texas. Conclude that health care providers respond differently to Medicare incentives compared to private insurance.

14. Gawande A. The cost conundrum.New Yorker [serial on the Internet]. 2009 Jun 1 [cited 2010 Nov 3].Available from: http://www.newyorker.com/reporting/2009/06/01/090601fa_fact_gawande

15. How do we test it? Education, Feedback, and Surveillance –If by providing education, or by monitoring certain types of treatments, there is a change  Practice Style Hypothesis. Some verification, but not a lot. Comparing Utilization Rates in Homogeneous Areas –If you can rule out utilization differences due to socioeconomic factors, you can say that practice style is important. Control by regression analysis. If you do a regression: Utilization =  b i x i + e, then if you’ve controlled for everything, you get an R 2 measure. Practice style would be the residual.

16. Three Problems How do you know if you’ve controlled for everything? What if some of your x’s actually represent practice style. Most of this is decidedly ad hoc. You’d like to see some good modeling. When done, we explain between 40 and 75% of the variation. This may leave a little, or a lot of variation to be explained by practice style.

17. SAV and Inappropriate Care Can you look at different levels of care, and determine that something wrong is going on. A> No! Efficient use of care occurs where marginal benefits = marginal costs. Simplest way to define this is to look at supply and demand curves. You may have single demand curve, and differing supply curves, due, e.g. to factor conditions Medical Care $ D1D1 S1S1 $ D1D1 S1S1 D2D2 S2S2 Demand? Supply?

18. SAV and Inappropriate Care OR, Differing demand curves due to incomes, preferences, et. Medical Care $ D1D1 S1S1 D2D2

19. + - Inappropriate Levels of Care If Q 1 is the “right” level, then What are the costs of being at the wrong level, Either + or -? Medical Care $ D1D1 S1S1 Q1Q1

20. Measuring the costs in the aggregate Fundamentally, we'll assume that marginal cost is constant. If marginal costs are rising, we'll see that these constitute lower bounds to the true costs. (1) W = 0.5   x i  v i Why? W = welfare loss x i = utilization of intervention for person i v i = valuation of incremental unit IF  is the correct utilization, (2) W = 0.5  (x i -  )  v i Loss triangles !

21. If we then multiply and divide (3) by  2, we get [write out]: W = 0.5 E  2 Nv/  W = (0.5 E  2 /  2 )(Nv/  )  2 Coefficient of variation =  / , so  2 /  2 is CV 2. Measuring the costs (2) Suppose that the valuation function is:  =  v/  x, or  v =  (x i -  ). Substituting into (2): (3) W = 0.5   (x i -  ) (x i -  ) = 0.5   2 N. Define inverse elasticity, at the mean, as: E = (dV/dx)(x/V) =   /v   = E v/ . x v xx vv Slope =  (2) W = 0.5  (x i -  )  vi xixi 

22. That leaves Nv , which equals aggregate spending. W = 0.5 * E * CV 2 * Spending Level. I like working in terms of real Elasticities, so I would use: W = 0.5 * CV 2 * Spending Level/E'. where E' is the true demand elasticity, and CV 2 is the coefficient of variation squared. Coefficient of variation is defined as the standard deviation divided by the mean. A good descriptive measure but it doesn't have a lot of statistical theory attached to it. Measuring the costs (3)

23. Measuring the costs (4) So this says that careful study of a medical intervention will have a greater expected benefit when: - large numbers of people are affected. - the per-unit cost of the intervention is high. - the level of uncertainty about correct use is large. - demand is inelastic. W = 0.5 * CV 2 * Spending Level/E'.

24. What if we use mean rather than X*? What may happen if we are comparing the actual use with the mean, when, instead, we should be using X*, the appropriate level? Consider valuation curves V 1 and V 2, their average V a, and the appropriate level V*. If we compute the welfare loss around average X a, we would include areas A, B, and C. The correct measure has areas A, C, D, E, and F, but not B. Measured Loss = A + B + C True Loss = A + D + E + F + C

25. B Rate of Use Incremental Value MC V* VA V2 V1 X1X1 X*X*XaXa X2X2 A C D E F Measured Loss = A + B + C True Loss = A + D + E + F + C True - Measured = (A + D + E + F + C) - (A + B + C) T - M = (D + E)

26. Measured Loss = A + B + C True Loss = A + D + E + F + C True - Measured = (A + D + E + F + C) - (A + B + C) T - M = (D + E) Region F has the same area as region B, so the missing area has size of regions D and E combined which is a parallelogram. So you are underestimating by (D + E). Of course, if marginal costs are increasing, the losses are even larger. What if we use mean rather than X*?