Evidence from Medicare How Risk Scores are Influenced by Insurers

Figure 1: How Risk Scores are Influenced by Insurers ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! Insurer'Actions' Contract!with! physician!on!a!risk! basis! Physician'Office' Actions' Encourage/require! Evaluation!Visit! Physician!Visit! Prepopulate! physician!notes! with!past!codes! Train!physician’s! staff!on!coding! Physician! takes!Notes! ! Billing!staff! codes!notes! ! ! Claims!and!codes! sent!to!Insurer! Dispatch!home! health!visit! ! Request!change!to! coding! No Directly! change! coding! ! Insurer! satisfied! with!codes?! Yes Final!codes! determine! risk!score! Note: The flowchart illustrates how diagnosis codes originate and how insurers can influence the process that generates them. Insurer actions are towards the left of the figure in blue boxes. Provider actions, including those of the provider’s billing and coding staff are towards the right in black boxes. Actions that immediately result in code data generation are represented by rhombuses. 33 Figure 2: Coding Across the Age 65 Threshold for Beneficiaries Choosing TM and MA Post 65 Pre 65 Post 65 Pre 65 Post 65 Pre 65 Post 65 .4 .25 .3 Any HCC Number of HCCs .5 .6 .7 Pre 65 (B) Count of Chronic Conditions .35 .4 (A) Coded with Any Chronic Condition Elected Traditional Medicare Elected Medicare Advantage Elected Traditional Medicare Elected Medicare Advantage (C) Risk Score Post 65 Pre 65 Post 65 .45 .5 Risk Score .55 .6 Pre 65 Elected Traditional Medicare Elected Medicare Advantage Note: Figure shows difference-in-differences summary statistics comparing the change in an individual’s risk score and diagnosed conditions at age 65 among consumers entering MA and consumers entering TM. Blue bars indicate individuals who choose to enroll in TM at age 65, and gray bars indicate individuals who choose to enroll in MA at age 65. Sample includes 4,724 Medigap enrollees and 1,347 MA enrollees. We describe the construction of these statistics in more detail in Appendix A.2. Figure 3: Identifying Coding Differences in Selection Markets 1.5 1.5 !! !! !! 1.0 1.0 Selection Selection !̅!̅ !! Average AverageRisk RiskScore Score !̅!̅ !̅!̅ !! 0.5 0.5 0.0 0.0 B B B Market Market share share // Penetration Penetration (θ (θB)) 1.0 1.0 1.5 1.5 !! !̅!̅ !! !̅!̅ !! !! 1.0 1.0 !̅ !! Selection Selection (B) Selection and Differential Coding Coding Coding !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! !! (A) Selection Only Average AverageRisk RiskScore Score !! !! !! !! !! !! !! !̅ !! !! !! !̅!̅!!!! !! !! !! 0.5 !! 0.5 !! !! 1.0 0.0 B 1.0 0.0 B B Market Market share share // Penetration Penetration (θ (θB)) !! !! !! Note: This figure demonstrates how to separate coding differences from selection when true underlying risk is !! B unobservable. The horizontal axis measures the market share of segment B, q . The vertical axis measures the average risk score: Average risk in A is r A , average risk in B is r B , and the average risk of all enrollees in the market is r. The dashed line in the figure represents the counterfactual average risk that segment B enrollees would have been assigned under segment A coding practices, r BA . All consumers choose either A or plan B. Segment B, which models Medicare Advantage, is assumed to be advantageously selected in both panels, and additionally is assumed to have higher coding intensity in the bottom panel. If and only if there are coding differences between A and B, ∂r then the slope of the market-level risk curve with respect to marketshare ( ∂q B ) will be different from zero. 35 Figure 4: Growth in Medicare Advantage (MA) Penetration 25 (A) National MA Penetration Changes Following MMA Implementation MA penetration, % 15 20 Part D Introduction 11 20 10 20 09 20 08 20 07 20 06 20 05 20 04 20 03 20 02 20 01 20 00 20 19 99 10 MMA (B) Within-county MA Growth -.15 -.1 -.05 0 .05 .1 .15 Penetration 2011 - Penetration 2006 .2 .25 .3 Note: The top panel displays national trends in MA penetration, where the unit of observation is the Medicare beneficiary. Source, pre 2006: Kaiser Family Foundation, 2013. The bottom panel displays a histogram of withincounty changes in penetration from 2006 to 2011, using the main estimation sample. The unit of observation is the county. Figure 5: Geography of Growth in Medicare Advantage (MA), 2006 to 2011 Quantiles (.13, .37] (.08, .13] (.04, .08] [-.25, .04] Note: Map shows changes by county in MA penetration from the beginning to end of our sample period, 2006 to 2011. Counties are binned and color-coded according to their quantile of changes in penetration. Darker regions indicate larger MA growth. ! ! ! Figure 6: Timing Illustration: Coding Effects Occur with a Lag in Medicare ! ! ! ! ! Coding differences ! can only enter Market Average Risk ! No immediate reported risk scores ! assuming new MA response in with a 1- or 2-year ! enrollees are new observed scores Coding effect lag. eligibles ! ! ! No immediate ! Market Average Risk response in observed scores ! assuming new MA Coding effect ! enrollees are switchers ! ! ! ! ! ! ! MA Penetration ! ! ! Year = t + 1 Year = t + 2 Year = t -1 Year = t ! Risk scores updated to Risk scores updated to MA advantageously Marginal enrollee i with ! include diagnoses include diagnoses selected. TM Avg > riski > MA ! entered between time t entered between time Avg, switches from TM ! and t+1 for enrollees t+1 and t+2. First year to MA or chooses MA enrolled all of t-1 coding can enter risk after becoming Medicare ! (switchers) scores for newly eligible ! eligible during year t ! ! ! Note: This diagram highlights the timing of changes in market level average risk in response to a change in MA ! For the first year in either MA or TM, a switcher carries forward a risk score based on his last year in penetration. the other !segment. For the newly eligible (those turning 65), demographic risk scores are assigned until there is a full year of enrollment and diagnoses information. Therefore, upcoding effects should not be apparent until a full year ! following! the change in enrollment when the penetration change is due to switchers, and upcoding effects should not be apparent until period t + 2 when the penetration change is due new Medicare enrollees. ! 38 Table 1: Summary Statistics Analysis)Sample:)Balanced)Panel)of)Counties,)2006)to)2011 2006 2011 Mean Std.)Dev. Mean Std.)Dev. Obs MA)penetration)(all)plan)types) Risk)(HMO/PPO))plans PFFS)plans Employer)MA)plans Other)MA)plans 7.1% 3.5% 2.7% 0.7% 0.2% 9.1% 7.3% 3.2% 2.2% 1.4% 16.2% 10.5% 2.7% 2.8% 0.0% 12.0% 10.5% 3.7% 4.3% 0.2% 3128 3128 3128 3128 3128 MALPart)D)Only)Penetration MA)nonLPart)D)Only)Penetration 6.5% 0.6% 9.5% 1.7% 13.1% 3.0% 10.8% 4.0% 3128 3128 Market)Risk)Score Risk)Score)in)TM Risk)Score)in)MA 1.057 1.064 0.949 0.084 0.087 0.181 1.054 1.057 1.032 0.090 0.089 0.155 3128 3128 3124 Ages)within)Medicare <65 65L69 70L74 75L79 ≥80 19.8% 23.5% 19.2% 15.9% 21.6% 6.3% 3.4% 1.9% 2.1% 4.4% 17.2% 23.7% 20.2% 15.4% 23.5% 6.2% 3.1% 2.5% 1.8% 5.0% 3128 3128 3128 3128 3128 Note: Table shows county-level summary statistics for the first and last year of the main analysis sample. The sample consists of of 3128 counties for which we have a balanced panel of data on penetration and risk scores. MA penetration in the first row is equal to the beneficiary-months spent in Medicare Advantage divided by the total number of total Medicare months spent in the county ⇥ year. The market risk score is averaged over all Medicare beneficiaries in the county. Table 2: Main Results Dependent(Variable:(County9Level(Average( Risk(Score (1) (2) (3) MA(penetration(t((placebo) 0.007 (0.015) 0.001 (0.019) 0.001 (0.019) MA(penetration(t91 0.069** (0.011) 0.067** (0.012) 0.064** (0.011) X X X X X X X X X 1.00 15,640 1.00 15,640 Main(Effects County(FE Year(FE Additional(Controls State(X(Year(Trend County9Year(Demographics Mean(of(Dep.(Var. Observations 1.00 15,640 Note: Table shows coefficients on contemporaneous (t) and lagged (t 1) Medicare Advantage (MA) penetration. The dependent variable is the average risk score in the market (county). Because MA risk scores are calculated using diagnosis data from the prior plan year, coding differences can plausibly affect risk scores only with a lag, and coefficients on contemporaneous penetration serve as a placebo test. Observations are county ⇥ years. All specifications include county and year fixed effects. Column 2 additionally controls for state indicators interacted with a linear time trend. Column 3 additionally controls for the demographic makeup of the county ⇥ year by including 18 indicator variables capturing the fraction of population in 5-year age bins from 0 to 85 and >85. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Table 3: Placebo Tests: Effects of Contemporaneous Penetration and Leads Dependent'Variable:'County1Level'Average'Risk'Score available'panel'years: 200712011 200712010 200712009 200812011 (1) (2) (3) (4) MA'penetration't+2'(placebo) 0.044+ (0.023) MA'penetration't+1'(placebo) 0.017 (0.025) 0.032 (0.056) MA'penetration't'(placebo) 0.001 (0.019) 10.021 (0.028) 10.064 (0.071) 0.006 (0.017) MA'penetration't11 0.064** (0.011) 0.076** (0.018) 0.084** (0.022) 0.041** (0.015) MA'penetration't12 Main'Effects County'FE Year'FE Additional'Controls State'X'Year'Trend County1Year'Demographics Observations 0.046* (0.022) X X X X X X X X X X X X X X X X 15,640 12,512 9,384 12,512 Note: Table shows coefficients on future (t + 2, t + 1), contemporaneous (t), and lagged (t 1, t 2) Medicare Advantage (MA) penetration. The dependent variable is the average risk score in the market (county). Because MA risk scores are calculated using diagnosis data from the prior plan year, coding differences can plausibly affect risk scores only with a lag, and coefficients on contemporaneous and future penetration serve as a placebo test. Observations are county ⇥ years. The data include penetration from 2006 through 2011 and market risk from 2007 through 2011. The inclusion of leads and lags determines the available panel years, listed in the header for each column. Controls are as described in Table 2. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Table 4: Falsification Test: Effects on the Demographic Portion of the Risk Score Dependent(Variable:(Demographic(Portion(of( County3Level(Average(Risk(Score (1) (2) (3) MA(penetration(t 0.000 (0.002) 0.001 (0.002) 0.001 (0.002) MA(penetration(t31 0.001 (0.002) 0.000 (0.002) 30.001 (0.002) X X X X X X X X X 0.485 15,640 0.485 15,640 Main(Effects County(FE Year(FE Additional(Controls State(X(Year(Trend County3Year(Demographics Mean(of(Dep.(Var. Observations 0.485 15,640 Note: Table shows coefficients on contemporaneous (t) and lagged (t 1) Medicare Advantage (MA) penetration. The dependent variable is the average demographic risk score in the market (county). Demographic risk scores are calculated by the authors using data from Medicare Beneficiary Summary File. Observations are county ⇥ years. Controls are as described in Table 2. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Table 5: Falsification Test: Effects on Morbidity and Mortality (1) Dependent+Variable: Mortality+over+65 Cancer+Incidence+over+65+ (2) (3) (4) (5) (6) MA+penetration+t 40.002 (0.002) 0.002 (0.002) 0.002 (0.003) 40.005 (0.004) 40.005 (0.005) 40.005 (0.005) MA+penetration+t41 0.002 (0.002) 40.002 (0.002) 40.002 (0.002) 0.005 (0.004) 0.001 (0.004) 0.003 (0.005) X X X X X X X X X X X X X X X X X X 0.048 15,408 0.048 15,408 0.023 3,050 0.023 3,050 Main+Effects County+FE Year+FE Additional+Controls State+X+Year+Trend County4Year+Demographics Mean+of+Dep.+Var. Observations 0.048 15,408 0.023 3,050 Note: Table shows coefficients on contemporaneous (t) and lagged (t 1) Medicare Advantage (MA) penetration. The dependent variable in columns 1 through 3 is the morality rate, derived from the National Center for Health Statistics, among county residents 65. The dependent variable in columns 4 through 6 is the cancer incidence rate among county residents 65. Cancer incidence data come from the Surveillance, Epidemiology, and End Results (SEER) Program by the National Cancer Institute. The smaller sample size in columns 4 through 6 reflects the incomplete geographical coverage of SEER cancer incidence data. Observations are county ⇥ years. Controls are as described in Table 2. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Table 6: Falsification Test: Effects on Medicare Age Distribution Fraction1≥65 Conditional1on1≥65 75#79 80#84 (4) (5) (1) 65#69 (2) 70#74 (3) MA1penetration1t 0.003 #(0.004) 0.002 #(0.007) 0.008 #(0.007) #0.004 #(0.006) #0.001 #(0.006) #0.006+ #(0.003) MA1penetration1t#1 #0.004 #(0.004) #0.006 #(0.006) 0.019** #(0.006) #0.006 #(0.007) #0.003 #(0.006) #0.004 #(0.004) X X X X X X X X X X X X X X X X X X 15,640 15,640 15,640 15,640 15,640 15,640 Main1Effects County1FE Year1FE Additional1Controls State1X1Year1Trend County#Year1Demographics Observations 85+ (6) Note: Table shows coefficients on contemporaneous (t) and lagged (t 1) Medicare Advantage (MA) penetration. The dependent variable in column 1 is the fraction of the Medicare population 65. The dependent variables in columns 2 through 6 are the fraction of the Medicare population in age bins conditional on age 65. Observations are county ⇥ years. Controls are as described in Table 2. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Table 7: The Role of Insurer-Provider Integration: Effects by Contract Type Dependent)Variable:)County:Level)Average)Risk)Score (1) (2) (3) HMO)penetration t)(placebo) 0.011 (0.033) 0.035 (0.036) 0.026 (0.035) PPO)penetration t)(placebo) :0.006 (0.033) 0.011 (0.038) 0.012 (0.037) PFFS)penetration t)(placebo) :0.015 (0.032) 0.000 (0.037) 0.001 (0.036) Employer)MA)penetration t)(placebo) 0.002 (0.011) :0.022+ (0.013) :0.021 (0.013) HMO)penetration t:1 0.137** (0.027) 0.098** (0.028) 0.101** (0.028) PPO)penetration t:1 0.115** (0.026) 0.072* (0.028) 0.068* (0.028) PFFS)penetration t:1 0.048* (0.023) 0.063* (0.025) 0.058* (0.025) Employer)MA)penetration t:1 0.036** (0.010) 0.041** (0.012) 0.041** (0.012) X X X X X X X X X 15,640 15,640 Main)Effects County)FE Year)FE Additional)Controls State)X)Year)Trend County:Year)Demographics Observations 15,640 Note: Table shows coefficients on contemporaneous (t) and lagged (t 1) Medicare Advantage (MA) penetration, disaggregated by contract type. The dependent variable is the average risk score in the market (county). Because MA risk scores are calculated using diagnosis data from the prior plan year, coding differences can plausibly affect risk scores only with a lag, and coefficients on contemporaneous penetration serve as a placebo test. Observations are county ⇥ years. Regressions additionally control for penetration of all other contract types, so that the net change in penetration summed across contract types equals the “MA penetration” variable in Table 2. Other controls are as described in Table 2. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Appendix APPENDIX A.1 Background on MA Risk-Adjusted Payments Medicare Advantage (MA) insurance plans are given monthly capitated payments for each enrolled Medicare beneficiary. The bases of these county-level payments are tied to historical Traditional Medicare (TM) costs in the county. County-based payments were originally intended to capture the cost of enrolling the “national average beneficiary” in the Traditional Medicare program in the county, though Congress has made many ad-hoc adjustments over time. Before 2004, there was relatively minimal risk adjustment of capitation payments, which relied primarily on demographics. In 2004, CMS began transitioning to risk adjustment based on diagnoses obtained during inpatient hospital stays and outpatient encounters. By 2007, diagnosis-based risk adjustment was fully phased-in. During our study period (2006-2011), risk adjusted capitation payj j ments were equal to Ric = fc · xic L, where i indexes beneficiaries, j indexes plans, and c indexes j counties (markets). The basis f j was approximately equal to the county “benchmark” fc , though fc could vary across plans within the same county. f could vary within counties because since 2006, MA plans have been required to submit bids to CMS, which are compared to the uniform county benchmark fc . If the bid is below the county benchmark set by the regulator, the plan receives 75% of the difference between the bid and benchmark, which the plan folds back into its premium and benefits as a “rebate” to beneficiaries. Importantly for our purposes, this 75% is still paid out by CMS into the MA program, so that what matters for the implicit subsidies to MA that we calculate here is the capitation payment to plans inclusive of any “rebate.” Therefore, we abstract from discussion of the details of how the rebate is allocated. A.2 Data Used in Section 2 For the analysis in Section 2, we obtained data on Medicare Advantage and Medigap claims from the new Massachusetts All-Payer Dataset (Mass APCD) from the Massachusetts Center for Health Information and Analysis (CHIA). The Mass APCD includes the universe of health insurance claims for individuals receiving medical services in the state of Massachusetts. Payers, along with third-party claims administrators and pharmacy benefit managers, report all claims to the state of Massachusetts. These claims are then aggregated into a large, comprehensive dataset. To identify individuals, we use an individual ID created by the state generated using Social Security numbers. This ID is available only for 2011-12. We identify two groups of individuals in the Mass APCD. The first group consists of all individuals enrolling in a product identified as a Medicare Advantage plan within 1 month of their 65th birthday. We identify Medicare Advantage plans using an identifier provided by CHIA. We verify that the products are Medicare Advantage products by matching the names of the payers in the Mass APCD data to publicly available Medicare Advantage enrollment data provided by CMS and by observing the age distribution of enrollees in these plans, focusing on the portion of enrollees over age 65. The second group consists of all individuals enrolling in a product identified as a Medigap plan within 1 month of their 65th birthday. We also identify Medigap plans using an identifier provided by CHIA. Again, we verify that the products are Medigap products by matching the names of the payers in the Mass APCD data to publicly available information about the insurers competing in the Massachusetts Medigap market. Additionally, we observe the portion of total spending paid by Medicare. For almost all of the plans identified as Medigap plans, this value was between 70 and 90%, confirming that these products most likely are Medigap plans. Appendix For both groups, we identify the subset of individuals with some form of coverage for every month of 2011 and 2012. We drop any individuals with some form of Medicare coverage prior to joining the products we identified as MA and Medigap products. Because Medicaid data is excluded from our version of the Mass APCD, this results in a sample of individuals with continuous commercial coverage prior to their 65th birthday and continuous Medigap/MA coverage after their 65th birthday. To ensure that we have enough data to calculate risk scores, we restrict the sample to individuals with at least 6 months of data prior to joining Medigap/MA and 6 months of data after joining Medigap/MA. This effectively eliminates anyone joining Medigap/MA between January and June 2011 and between August and December 2012. We retrieve the claims for each individual in the sample. We restrict the claims sample to claims from medical plans only, excluding prescription drug plans.50 We then assign individuals to cohorts based on the month in which they joined Medigap/MA. We specify separate “pre-65” and “post65” periods for each cohort. We require that the pre-65 and post-65 periods be of equal length and that they consist of the same months from different years. Given those restrictions, we choose the longest possible period for each cohort. For example, for the individuals joining Medigap/MA in July 2011, we specify the pre-65 period to be January-June 2011 and the post-65 period to be JanuaryJune 2012. For the individuals joining Medigap/MA in February 2012, we specify the pre-65 period to be February-December 2011 and the post-65 period to be February-December 2012.51 For each individual, we calculate a pre-65 and a post-65 risk score. The pre-65 (post-65) risk score is calculated by identifying all diagnoses from medical claims incurred during the pre-65 (post65) period and running those diagnoses through the 2011 version of the CMS-HCC SAS program provided by CMS. The program maps each individuals diagnoses to a set of 70 chronic conditions. It then multiplies indicators for these chronic conditions by a set of weights estimated using claims data from Traditional Medicare. The product of the individual’s chronic conditions indicators and the weights is the risk score. We also use the chronic conditions indicators to calculate each individual’s number of chronic conditions and to generate an dummy variable indicating whether the individual has at least one chronic condition. To produce the statistics reported in Section 2, we adjust the risk scores for gender and the region of Massachusetts in which the individual resides. We also adjust for cohort fixed effects to account for the fact that the risk scores are calculated using time periods of different lengths. We do the adjustment by running a regression of risk scores on these variables and extracting the residuals. The adjusted risk scores are equal to the residuals plus the estimated intercept from the regression. Using individual fixed effects produces similar results. A.3 Linearity of Market-level Average Risk Curve Throughout Section 3 we impose the assumption that coding differences can be represented by an additive factor. Under this assumption, the slope of the average risk curve in the market is exactly equal to the upcoding factor. It is true, however, that the weaker assumption that any individualspecific heterogeneity in the plan coding factor is orthogonal to Segment B’s market share leads to the similar result: The slope of the average risk curve in the market is exactly equal to the average upcoding factor across individuals in the market. To formalize this statement, let individual i’s risk score in Plan A be equal to riA = r̂i + a A , and let i’s risk score in Plan B be equal to riB = r̂i + a B + eiB where eiB has mean e and represents 50 The CMS-HCC risk adjustment model used for MA payments is based only on diagnoses from medical claims. we could have required that the pre-65 and post-65 periods be contiguous. For example, for individuals joining Medigap/MA in July 2011, we could specify the pre-65 period to be January-June 2011 and the post-65 period to be July-December 2011. This alternative specification produces results nearly identical to those produced by the specification outlined above. 51 Alternatively, Appendix individual-specific heterogeneity in the relative coding intensity between plans A and B. For individual i, this generates an “upcoding factor” of a B + eiB a A . This produces an average upcoding 1 factor of (a B + eiB a A )) = a B + e a A . NÂ Letting 1 [ Bi (q )] represent the indicator function for choosing B, market average risk can be 1 expressed as r = (r̂i + a A + 1 [ Bi (q )](a B + eiB a A )). Assuming that ei is orthogonal to q, NÂ ∂r ∂ 1 = differentiating r with respect to q produces, (r̂i + a A + 1 [ Bi (q )](a B + eiB a A )) = ∂q ∂q N Â ∂r a B + e a A . Since e is constant, â B can be defined as a B + e so that = â B a A , the average ∂q upcoding factor. A.4 Estimates of Selection Section 5 describes the results of the main analysis in which we regress county-level averages of risk scores on lagged MA penetration in the county to estimate coding differences. Here we estimate selection on risk scores, using an analogous set of regressions. Under the assumption that MA penetration changes are exogenous to changes in underlying population health conditional on our controls, selection can be estimated by regressing either the average county risk score within TM or the average county risk score within MA on contemporaneous and lagged penetration. Table A2 presents the selection results. Coefficients on contemporaneous MA penetration identify pure selection effects. Positive contemporaneous coefficients in both markets would indicate that as penetration increased, the marginal beneficiary choosing MA was high risk relative to the MA average and low risk relative to the TM average, increasing the average risk score in both pools. In Table A2, estimates for both FFS and MA risk are imprecise, yielding confidence intervals consistent with a broad range of selection effects, including the findings in Newhouse et al. (2012) of advantageous selection into MA of 4 to 9% of the risk score in 2008.52 An important component of selection effects may be captured by the lagged penetration coefficient: Research on MA enrollment by Sinaiko, Afendulis and Frank (2013) shows that the majority of new MA enrollees are age 65, implying that most of the shift in MA penetration is likely occurring among the newly Medicare-eligible. In Table A2, this would cause a significant fraction of selection effects to be captured by the lagged coefficient, as new MA enrollees aren’t assigned risk scores until their second year. In short, interpreting selection effects in Table A2 is difficult because of the differential riskscoring treatment of the newly-eligible 65 year-olds. Coefficients on lagged MA penetration are affected by: (i) selection on risk score trajectory and (ii) selection on unobserved contemporaneous risk score for new enrollees who are not assigned a diagnosis-based score until their second year. It is important to note that unlike these within-market-segment results, the regressions comprising our main analysis, which examine lagged coefficients on overall county risk, are unaffected by selection and yield a more straightforward identification of pure coding effects. 52 Regressions in which the dependent variable is the risk score estimate only compensated selection. This is in contrast, for example, to Brown et al. (2014) and Cabral, Geruso and Mahoney (2014) which are primarily interested in uncompensated selection. !̅ ! 1.0 !̅ ! 0.5 ! ! ! 1.0 Selection Coding !̅!! Average Risk Score Selection !̅ ! 1.5 !̅ !̅ ! 0.5 0.0 B Market share / Penetration (θB) 1.0 (D) Selection is Nonlinear and Non-monotic 1.5 !̅ ! !̅ ! ! 1.0 Coding ! ! 1.0 Coding Average Risk Score !̅ ! 1.0 (B) Advantageously Selected Plan Codes Less Intensely Average Risk Score 1.5 Selection ing ! ! !̅ ! ! !̅ ! ! ! !̅ ! ! ! ! ! !̅!! ! ! ! ! 0.5 0.5 ! ! ! ! ! ! 1.0 1.0 0.0 0.0 B Market share / Penetration (θB) B Market share / Penetration (θB) ! ! ! ! ! ! ! when true underlying risk is unobservable. The horizontal axis ! Note: This figure demonstrates how to separate coding differences from selection ! ! B . The vertical axis measures the average measures the market share of segment B, q risk score: Average risk in A is r A , average risk in B is r B , ! ! and the ! average risk of all enrollees in the !market is r. The dashed line in the figure represents the counterfactual average risk that segment B !̅ 1.5 enrollees! would have been assigned under segment A coding practices, r BA . All consumers choose either A or plan B. If and only if there are coding ! ∂r differences between A and B, then the slope !̅!!of the market-level risk curve with respect to marketshare ( ∂q B ) will be different from zero. ! ! ! !̅ Score 49 ! ! ! ! ! ! Coding Average Risk Score Figure A1: Identifying Coding Differences 0.0 in Selection Markets B Market share / Penetration (θB) ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! C Average R Average Ri ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! Appendix ! ! 1.0 ! ! ! !̅ ! ! ! ! 0.5 ! ! ! 1.0 0.0 B Market share / Penetration (θB) ! ! ! (A) Advantageously Selected Plan Codes More Intensely ! ! ! ! ! 1.5 ! !̅ ! ! ! !̅ ! ! ! ! 1.0 !̅ ! ! ! ! !̅!! ! ! ! 0.5 ! ! ! 1.0 0.0 B Market share / Penetration (θB) ! ! ! ! (C) No Selection and Differential Coding Intensity Appendix Table A1: Difference-in-Differences: Coding Across the Age 65 Threshold for Beneficiaries Choosing TM and MA Risk7Score (1) (2) Post-65 Selected7MA Post-657X7Selected7MA 0.023* (0.011) -0.072** (0.013) 0.044* (0.021) Person7FE Mean7of7Dep.7Var. Observations 0.023** 0.007 Dependent7Variable:7 Count7of7HCCs (3) (4) 0.031 (0.023) -0.175** (0.027) 0.044** 0.015 0.123** (0.043) X 0.55 12,142 0.031* (0.015) 0.55 12,142 At7least717HCC (5) (6) 0.005 (0.010) -0.071** (0.014) 0.123** (0.030) 0.060** (0.020) X 0.57 12,142 0.005 (0.007) 0.57 12,142 0.060** (0.015) X 0.33 12,142 0.33 12,142 Note: Table shows coefficients from difference-in-differences regressions that compare the change in coded diagnoses across the age 65 threshold among consumers who select TM relative to consumers who select MA. The dependent variables are the risk score in columns 1 and 2, the count of HCCs in columns 3 and 4, and an indicator for being diagnosed with a condition yielding at least one HCC in columns 5 and 6. Odd columns include individual fixed effects. Data are as described in Section A.2. Observations are person-years. Standard errors in parentheses are clustered at the person level. + p < 0.1, * p < 0.05, ** p < 0.01. Dependent+Variable: Mean+TM+Risk+Score Mean+MA+Risk+Score (1) (2) (3) (4) (5) (6) MA+penetration+t 0.037 (0.026) 0.040 (0.034) 0.040 (0.033) 0.025 (0.062) 70.024 (0.085) 70.013 (0.083) MA+penetration+t71 0.045** (0.013) 0.030* (0.012) 0.026* (0.012) 0.087* (0.040) 0.116** (0.040) 0.130** (0.041) X X X X X X X X X X X X X X X X X X 1.006 15,640 1.006 15,640 0.959 15,616 0.959 15,616 Main+Effects County+FE Year+FE Additional+Controls State+X+Year+Trend County7Year+Demographics Dep+var+mean Observations 1.006 15,640 0.959 15,616 Note: Table shows coefficients on contemporaneous (t) and lagged (t 1) Medicare Advantage (MA) penetration. The dependent variable in columns 1 through 3 is the average risk score among TM enrollees in the market (county). The dependent variable in columns 4 through 6 is the average risk score among MA enrollees in the market (county). Observations are county ⇥ years. Controls are as described in Table 2. Standard errors in parentheses are clustered at the county level. + p < 0.1, * p < 0.05, ** p < 0.01. Appendix Table A2: Selection Results: Effects on within-TM and within-MA Risk Scores Appendix Table A3: Robustness to Alternative Subsets of the Identifying Variation Dependent.Variable:.County