Appendix. We used matched-pair cluster-randomization to assign the. twenty-eight towns to intervention and control. Each cluster,

Transcription

1 Yip W, Powell-Jackson T, Chen W, Hu M, Fe E, Hu M, et al. Capitation combined with payfor-performance improves antibiotic prescribing practices in rural China. Health Aff (Millwood). 2014;33(3). Published online February 26, Appendix Experimental design and randomization We used matched-pair cluster-randomization to assign the twenty-eight towns to intervention and control. Each cluster, a town, consists of a THC, all the VPs under the THC s supervision and the population in its catchment area. On average, each town had 11 VPs and each VP served on average 1,500 people. The majority of VPs were staffed by one village doctor. We first paired towns before randomization, ensuring that matches were as similar as possible on a range of baseline characteristics (whether the THC provided hospitalisation services; whether the THC was classified as a centre THC to indicate its higher level of capacity; distance to the county seat; average expenditure per outpatient visit; number of outpatient visits per year; number of village posts under its management and percentage of agricultural population in the town to indicate socioeconomic development). 1,2 We then randomly assigned one cluster from each pair with the flip of a coin to receive the provider payment intervention described above, starting July 2010, and the other in the pair as control. All towns agreed to their assignment. However, a small-sized control THC lost its manager after the intervention began. The

2 county health bureau requested an intervention THC (not in the same matched pair) to manage this THC, which it did as if it were subject to capitation plus pay-for-performance incentives. We therefore dropped the paired cluster to which this control THC belonged. These events could not have been anticipated and serve to illustrate the politically robust usefulness of the pair-cluster randomised deign in this setting. 2 At the time of writing, the payment intervention in the intervention clusters remains ongoing. Identical trainings on appropriate drug prescription were provided to both the intervention and control THCs and VPs. THCs and their VPs were not masked to the intervention. Cluster residents, however, were masked in the sense that there was no public announcement of the change in payment system. All residents faced the same NCMS insurance policies and other relevant policies introduced in Ningxia province. We anticipated heterogeneity in intervention THCs compliance with the policy intervention, and tried to minimise this risk by having teachers from Ningxia Medical University conduct monitoring visits to the NCMS offices, THCs and selected VPs on a monthly but random basis. To examine whether there are compositional changes in patients due to the intervention, for example, changes in demand when patients respond to the capitation plus p4p payment

3 intervention by selecting into the control health centres/village posts, which might therefore bias the results, we use the household data to show that the characteristics of patients seeking care are very similar between intervention and control. These results are shown below in Exhibit A10. This is not surprising because from qualitative interviews with households, they were not aware of what provider payment methods their health centres or village posts faced. This is also consistent with the results shows in Exhibit 9A which showed that there was no effect of the intervention on utilization of healthcare. We further checked and found that there was very little movement between intervention and control clusters in the health seeking behaviour of residents. Only 4.8% of patient visits involved individuals living in an intervention (or control) cluster and seeking care from a control (or intervention) cluster. Validating randomization Using data from a survey of households and a survey of health providers conducted in 2009, we validate randomization by comparing the treatment and control groups. Characteristics of individuals living within study clusters assessed at baseline in the household survey were found to be similar between intervention groups, with the exception of ethnicity (Exhibit A3). Educational attainment and per capita consumption are low by Chinese standards. THC and VP characteristics also appeared

4 similar between groups (Exhibit A4). There were no notable differences on a wide range of variables measured at baseline in the health provider surveys. Sample size We conducted the sample size estimation separately for THCs and VPs on the basis of 28 clusters. The sample size at the level of THC assumed 49% of patients are prescribed antibiotics at baseline, with 5000 patient visits per cluster and a coefficient of variation of 0 2 providing 80% power at a 5% level of significance to detect a 15% (7 5 point) fall due to the intervention. 3 For VCs, we assumed 38% of patients are prescribed antibiotics at baseline, with 4000 patient visits cluster and a coefficient of variation of 0 2 providing 80% power to detect a 15% (5 7 point) treatment effect. Empirical analysis The effect of the intervention was estimated by fitting regressions of each outcome on a binary indicator of treatment status. For continuous outcomes, we used least squares regressions of the form: where is the outcome observed for individual in

5 village/township, is a binary indicator taking value 1 if village/township belongs to the treatment group (and 0 otherwise), is a vector of individual characteristics including gender and patient age, captures unobserved paired township characteristics through pair fixed-effects and is an independent error term such that. In the case of binary outcomes, we used a logistic regression of the form: and report marginal effects. The expenditure data showed a relatively small number of implausible values. We therefore trimmed the sample at the th percentile, while noting that estimates on the full sample are almost identical. We report unadjusted estimates (excluding and ) as well as estimates adjusted for the inclusion of patient gender, patient age and a dummy variable for each pair of matched clusters which accounted for heterogeneity across pairs of clusters. Robust standard errors, clustered at town level, were computed to allow for arbitrary correlations of observations within clusters. 4 For all outcomes, except patient satisfaction, we conducted a subgroup analysis by patient gender. For antibiotic use, we reported results limiting the

6 sample to patients diagnosed with a cold to provide more clear-cut evidence on the extent of unnecessary prescribing practices. All statistical analyses were done with Stata (version 12). Endnotes 1. King G, Gakidou E, Imai K, Lakin J, Moore RT, Nall C, et al. Public policy for the poor? A randomised assessment of the Mexican universal health insurance programme. Lancet. 2009; 373(9673): King G, Gakidou E, Ravishankar N, Moore RT, Lakin J, Vargas M, et al. A "Politically Robust" Experimental Design for Public Policy Evaluation, with Application to the Mexican Universal Health Insurance Program. J Policy Anal Manag. 2007;26: Hayes RJ, Bennett S. Simple sample size calculation for cluster-randomized trials. Int J Epidemiol. 1999; 28(2): Moulton BR. An Illustration of a Pitfall in Estimating the Effects of Aggregate Variables on Micro Units. Rev Econ Statistics. 1990; 72(2):

7 Exhibits Exhibit A1. Indicators of performance and scoring system under capitation with pay-for-performance Antibiotic prescription IV-antibiotic prescription Item Item Description Weight Score Data used for calculating performance score Falsifying visits Completeness of visit record Patient satisfaction Percentage of visits prescribed with antibiotics 150 Weight x (1 - % visits with antibiotics prescribed) Among visits with antibiotics prescribed, percentage given via IV injection Percentage of visits with identical patient name and health problems repeated within one day Percentage of visit record that has at least one of the following incomplete: symptoms; diagnosis; drugs prescribed (types, dosage); exam/test prescribed (for township health centres); expenditure Are you satisfied with the cleanliness of the clinic (1) Very satisfied (2) so- so (3) not satisfied Are the providers patient and careful in explaining to you your health problems? (1) Always (2) sometimes (3) never Do the providers explain to you how to take your medications? (1) Always (2) sometimes (3) never Are you satisfied with the providers technical quality? (1) Definitely (2) so- so (3) no Are you able to see the provider without long waiting? (1) Always (2) sometimes (3) never TOTAL Weight x (1 - % with IV injection among those prescribed antibiotics) 50 Weight x (1 - % of visits with identical patient name and health problems repeated within one day) 50 Weight x (1 - % visit record deemed incomplete) Management information system Management information system Management information system Random sample of 200 records 20 Weight x (% response 1 ) Interview with 30 randomly selected households in each 20 Weight x (% response 1 ) village 20 Weight x (% response 1 ) 20 Weight x (% response 1 ) 20 Weight x (% response 1 )

8 Exhibit A2. Description of the data sources used in the study Source of data Outcomes Description of survey Management information system Antibiotic use per visit Expenditure per visit Patient visits per day Study information system providing data on the universe of patient visits across a census of township health centres in the study area. Household survey Village clinic survey Patient satisfaction across the following dimensions: waiting time; cleanliness; doctor politeness; doctor explanation of illness and treatment plan; drug availability; equipment; and confidence in doctor. Data also used to assess differences in baseline characteristics between treatment and control Data used to assess differences in baseline characteristics between treatment and control. Two-stage sample survey representative of each town within a county. In each study town (ie. cluster), 40 percent of villages were randomly selected, and within each village 33 households were randomly interviewed. In 2009, 3,828 households were interviewed. In 2012, 3,887 households were interviewed. Random sample of village clinics, covering 40 percent of all village clinics in the study area. In 2009, 114 village clinics were interviewed. Township health centre survey Data used to assess differences in baseline characteristics between treatment and control. Census of all township health centres in the study area. In 2009, 26 township health centres were interviewed.

9 Exhibit A3. Household characteristics at baseline Total number of individuals Intervention (capitation 7989 with P4P) Control (fee-for-service) 8877 Age (years) 31.4 (19.4) 30.3 (19.4) Male 4170/7989 (52%) 4601/8877 (52%) Han 4861/7989 (61%) 4176/8877 (47%) Hui 3092/7989 (39%) 4675/8877 (53%) Other ethnicity 36/7989 (0.4%) 26/8877 (0.3%) No education 2340/7989 (29%) 2622/8877 (30%) Elementary school 2876/7989 (36%) 3540/8877 (40%) Middle school 2016/7989 (25%) 2072/8877 (23%) High school or above 757/7989 (9%) 643/8877 (7%) Migrant worker 1457/7989 (18%) 1574/8877 (18%) Consumption per capita 5090 (5271) 5409 (9469) Distance village clinic (km) 3.9 (13.6) 5.5 (27.9) Distance town health centre (km) 19.9 (16.4) 20.4 (16.9) Distance county hospital (km) (66.5) 82.4 (56.9) Head of the household 1848/7989 (23%) 1980/8877 (22%) Female headed household 307/7989 (4%) 327/8877 (4%) Household size (members) 5.0 (1.4) 4.8 (1.4) Source: Authors analysis of study data Notes: Data are n/n (%) or mean (standard deviation) unless otherwise stated. Data from the household survey conducted in 2009, before the intervention started, in the study clusters.

10 Exhibit A4. Health provider characteristics at baseline Township health centre Intervention (capitation with P4P) Control (fee-for-service) Number of physicians 18.1 (16.7) 20 (15.3) Number of nurses 10.2 (6.1) 10.7 (6.8) Number of other medical staff 3.6 (5.1) 4.5 (6.6) Number of beds 13.3 (19.4) 11.9 (9.6) Inpatient admissions per year (497.4) (527) Outpatient visits per year (12955) (6317) Income per year ( ) (666706) Cost per year ( ) (612418) Service revenue per year (253446) (244136) Village clinic doctor Doctor's years of experience 23.2 (14.2) 23.9 (15) Barefoot indicator 34/54 (62.96%) 40/60 (66.67%) Doctor has multiple jobs 40/54 (74.07%) 42/60 (70%) Male 46/54 (85.19%) 11/60 (81.67%) Age 45.8 (12.3) 45.9 (13) Doctor's household size 5 (1.9) 4.8 (1.5) Source: Authors analysis of study data Notes: Data are n/n (%) or mean (standard deviation) unless otherwise stated. Data from the health provider surveys conducted in 2009, before the intervention started, in the study clusters.

11 Exhibit A5. Effect of payment intervention on antibiotic prescribing practices (full version of Exhibit 3) Control mean Antibiotic use in township health centres Unadjusted Adjusted Treatment effect Treatment effect p value (95% CI) (95% CI) All (-0.211, ) (-0.123, ) Oral antibiotics (-0.110, 0.025) (-0.057, 0.028) Injectable antibiotics (-0.168, 0.013) (-0.104, 0.002) Patient diagnosed with a cold (-0.239, 0.140) (-0.173, ) Male (-0.200, 0.007) (-0.118, ) Female (-0.223, ) (-0.130, ) Antibiotic use in village posts All (-0.132, 0.027) (-0.115, ) Oral antibiotics (-0.091, 0.041) (-0.068, 0.014) Injectable antibiotics ( 0.000, 0.049) (-0.072, ) Patient diagnosed with a cold (-0.263, 0.004) (-0.245, ) Male (-0.138, 0.018) (-0.136, ) Female (-0.125, 0.038) (-0.105, 0.001) Adjusted estimates include cluster pair fixed effects as well as controls for patient gender and patient age. There are 440,473 patient visits (208,482 treatment group; 231,991 control group) in the township health centre analysis and 714,661 patient visits (338,185 treatment group; 376,476 control group) in the village post analysis. p value

12 Exhibit A6. Effect of payment intervention on expenditure per visit (full version of Exhibit 4) Control mean Unadjusted Total expenditure per visit in township health centres Adjusted Treatment effect Treatment effect p value (95% CI) (95% CI) p value All (-5.56, 4.66) (-5.48, 5.52) Male (-5.27, 4.34) (-5.01, 4.92) Female (-5.87, 5.08) (-5.99, 6.22) Total expenditure per visit in village posts All (-1.27, 0.34) (-1.65, -0.42) Male (-1.24, 0.42) (-1.62, -0.39) Female (-1.33, 0.31) (-1.70, -0.44) Drug expenditure per visit in township health centres All (-4.67, 2.54) (-4.28, 2.52) Male (-4.74, 2.63) (-4.19, 2.61) Female (-4.63, 2.50) (-4.39, 2.45) Drug expenditure per visit in village posts All (-0.57, 0.77) (-0.64, 0.16) Male (-0.55, 0.84) (-0.60, 0.17) Female (-0.61, 0.73) (-0.698, 0.17) Adjusted estimates include cluster pair fixed effects as well as controls for patient gender and patient age. Total expenditure and drug expenditure are trimmed at the percentile. There are 440,144 observations (208,300 treatment group; 231,844 control group) in the township health centre trimmed sample and 714,304 observations (338,031treatment group; 376,273 control group) in the village post trimmed sample. Total expenditure includes spending on drugs, visit fee, tests and diagnostics.

13 Exhibit A7. Distribution of total expenditure per visit in township health centres Density Total expenditure per visit Intervention Control Exhibit A8. Distribution of total expenditure per visit in village posts Density Total expenditure per visit Intervention Control

14 Exhibit A9. Effect of payment intervention on healthcare utilisation and patient satisfaction Control mean Unadjusted Number of patient visits per day in township health centres Adjusted Treatment effect Treatment effect p value (95% CI) (95% CI) p value All (-27.6, 19.0) (-19.59, 8.95) Male (-12.89, 9.58) (-9.03, 4.30) Female (-15.45, 9.96) (-11.12, 4.39) Number of patient visits per day in village posts All (-5.64, 2.49) (-3.60, 1.80) Male (-3.14, 1.34) (-1.94, 0.94) Female (-2.91, 0.99) (-1.85, 0.67) Patient satisfaction score Township health centres (-0.88, 1.19) (-0.69, 0.62) Village posts (-0.84, 0.60) (-0.63, 0.43) Adjusted estimates include cluster pair fixed effects as well as controls for patient gender and patient age. For healthcare utilisation, the unit of observation is the number of visits per day in each health provider. For patient satisfaction, the unit of observation is a household, irrespective of whether they have used health services. The patient satisfaction score ranges between 7 and 35.

15 Exhibit A10: Characteristics of patients seeking care during the intervention period intervention control #observation mean #observation mean t-test (cluster at town level) sought care in village posts (2012) age male circulatory system diseases respiratory system diseases digestive system diseases genitourinary system diseases musculoskeletal system diseases other diseases (not in top 5) chronic disease 163 sought care in township health centres (2012) age male circulatory system diseases respiratory system diseases digestive system diseases genitourinary system diseases musculoskeletal system diseases other diseases (not in top 5) chronic disease

16 sought care in village posts or township health centres (2012) age male circulatory system diseases respiratory system diseases digestive system diseases genitourinary system diseases musculoskeletal system diseases other diseases (not in top 5) chronic disease Data is from household survey data Patients are those who sought outpatient care in the past two weeks dating back from the day of interview.

17 Exhibit A11: Non-incentivized maternal health indicators (household data) intervention control #observation mean #observation mean t-test (cluster at town level) gynaecological exam deliver baby #prenatal exam measure weight in prenatal exam blood test in prenatal exam measure blood pressure in prenatal exam urine exam in prenatal exam have ultrasound in prenatal exam Data is from household survey Sample is female aging and on site during the interview. Except for the first two indicators, the other indicators were only asked for women who had a birth in the year preceding the survey.

18 Exhibit A12: Non-incentivized child health indicators (household data) #observation intervention mean #observation control mean t-test (cluster at town level) #check-up #DPT #polio #hepatitisb BCG measles meningitis encephalitis hepatitisa MMR Data is from household survey Sample is children 7 years old or younger.