Table 5.

Conventional multivariate regression, instrumental variable, and propensity score estimates of risk differences of hospital admission for self-harm (Hospital Episode Statistics data) or death by suicide (Office of National Statistics mortality data) within 3 months of index prescription (number of patients = 394,846^a and number of physicians = 3,042)

	TCAs vs. SSRIs risk differences (95% CI)
	Ordinary least squares regression	Instrumental variable analysis using one prior prescription	Instrumental variable analysis using one prior prescription and physician fixed effects	Instrumental variable analysis using seven prior prescriptions	Propensity score adjustmentb
TCA (reference category SSRIs)	−0.11 (−0.14, −0.08)	−0.04 (−0.21, 0.13)	−0.06 (−0.58, 0.45)	−0.10 (−0.20, 0.01)	−0.05 (−0.08, −0.03)
F-test		1,727	596
Hausman test (P-value)		0.45	0.83

Abbreviations: TCA, tricyclic antidepressant; SSRI, selective serotonin reuptake inhibitor; CI, confidence interval; BMI, body mass index.

CIs allow for clustering by physician. Patients previously admitted to hospital for self-harm are omitted. Reported F-statistic is robust and is a test of the partial association of the instrument and the prescription. The null hypothesis of the Hausman test is that there is no difference between the conventional ordinary least squares estimates and the instrumental variable results.

^aBased on the 50% of practices that were linked to the Hospital Episodes Statistics and ONS databases.

^bN = 394,836 because one patient with missing gender was excluded and nine patients were not in the common support for the propensity score adjustment results. Propensity score based on all covariates in Table 1 except BMI because BMI had 86,488 missing values.