Effects of aggregation of drug and diagnostic codes on the performance of the high-dimensional propensity score algorithm: an empirical example

Table 3 Risk ratios for different cohort size, outcome incidence and exposure prevalence of initiators of celecoxib or NSAIDs (ibuprofen or diclofenac) in a cohort 18–65 years old between 1 July 2003 and 30 September 2004 in the MarketScan database by using the High-Dimensional Propensity Score (hd-PS) adjustment with different aggregation methods

Cohort and variable selection method		Base scenario	Medications						Medical diagnoses								Combined
			No Rx	ATC* Level					No Dx	CCS^†Level					ICD-9^‡		ATC 4th + CCS 1st
			No Rx	1st	2nd	3rd	4th	5th	No Dx	1st	2nd	3rd	4th	Universal	3-digit	4-digit	ATC 4th + CCS 1st
Unadjusted		1.05
Basic covariates		0.98
Basic and extended covariates		0.95
Basic and hd-PS covariates		0.92	0.94	0.93	0.92	0.92	0.90	0.91	0.88	0.90	0.89	0.92	0.92	0.94	0.95	0.94	0.85
	%^§		3.9	2.6	0.0	0.8	−2.9	−1.4	−7.0	−3.7	−4.4	0.10	1.0	3.6	5.1	4.1	−12.1
Basic, extended and hd PS covariates		0.94	0.91	0.96	0.94	0.94	0.90	0.93	0.91	0.91	0.92	0.95	0.94	0.96	0.96	0.95	0.88
	%^§		−5.0	3.7	−0.5	−0.7	−6.0	−1.3	−5.0	−4.4	−2.5	1.0	0.6	3.6	4.0	2.1	−10.9
hd-PS covariates (k = 500)^║
Outpatient diagnoses (n)		136	224	198	177	154	144	133	0	32	90	97	54	123	133	139	34
Inpatient diagnoses (n)		9	12	11	11	9	9	7	0	22	18	19	5	16	14	11	23
Medication (n)		167	0	36	76	122	148	177	247	216	186	181	213	171	166	163	194
Outpatient procedures (n)		152	220	211	194	174	161	148	210	188	166	163	187	153	151	151	206
Inpatient procedures (n)		36	44	44	42	41	38	35	43	42	40	40	41	37	36	36	43

Abbreviations: basic covariates included continuous age, gender and calendar year; extended covariates included covariates adjusted for in published studies; hd-PS, high dimensional propensity score. Base scenario used up to 5-digit ICD-9, procedures, generic drugs for five data dimensions of the hd-PS.
No Rx: the scenario using up to 5-digit ICD-9 and procedures for 4 data dimensions of the hd-PS.
No Dx: scenario using procedures and generic drugs for 3 data dimensions of the hd-PS.
^*ATC: 5 levels of the Anatomical Therapeutic Chemical classification.
^†CCS: four levels of the Clinical Classification Software; Universal, the most granular CCS code available for each ICD-9 code.
^‡ ICD-9: International Classification of Diseases, 9th Revision, Clinical Modification.
^§: % proportional difference in absolute degree of estimated confounded between estimates for the specific aggregation scenario and the basic scenario at the same variable selection method on the natural log scale with RCT finding of 0.5. The presumptive amount of confounding in the basic scenario A = │ln(adjusted RR) – ln(0.5)│; in each aggregation method B = │ln(adjusted RR) – ln(0.5)│; and C = B/A –1.
^║: number of hd-PS covariates retained in the final propensity score model. Total (k=500) and from each data dimension (n).

ISSN: 1471-2288