Table 3.

Goodness of fit of alternative specifications of model for total patient cost.

Model specification	Log, Poisson				Linear, Gaussian (OLS)
	BIC	$R_D^2$	$R_{COR}^2$	MAE	BIC	$R_{COR}^2=R_D^2=R^2$	MAE
Age, gender (Model 1)	38446223	0.21	0.13	285	1320939	0.13	285
Age, gender, and deprivation (Model 2)	38041309	0.22	0.13	283	1320555	0.13	284
Age, gender, and practice (Model 3)	37725187	0.22	0.14	282	1322027	0.14	283
Age, gender, deprivation, and practice (Model 4)	37522639	0.23	0.14	281	1321883	0.14	282
(Model 4) + QOF disease indicators	29339460	0.40	0.25	244	1302791	0.31	234
(Model 4) + QOF chronic disease count	28523441	0.42	0.29	239	1305547	0.29	240
(Model 4) + Charlson indicators	32546370	0.33	0.22	259	1310235	0.25	255
(Model 4) + Charlson Index score	32274547	0.34	0.23	258	1311982	0.23	260
(Model 4) + EDC indicators	26255449	0.46	0.29	231	1295660	0.37	222
(Model 4) + EDC count	26196861	0.46	0.32	229	1302546	0.31	236
(Model 4) + ACG	28694630	0.41	0.27	242	1308944	0.27	248
(Model 4) + RUB	30367472	0.38	0.24	250	1312522	0.23	259

Notes. BIC: Bayesian Information Criterion. Smaller BIC indicates better fit and is comparable for different models with same error distribution. $R_D^2$: deviance based R², which is comparable for different models with same error distribution. $R_{COR}^2$: squared correlation coefficient from OLS regression of estimated cost on actual cost. For OLS models $R_{COR}^2$, the deviance based $R_D^2$, and the model R² are equal. MAE: mean absolute error. Estimation sample: 85,946 patients aged 20+ in 174 practices.