Bias in estimates of alcohol use among older people: selection effects due to design, health, and cohort replacement

Table 4 Sex-adjusted odds ratios for a more frequent^a alcohol consumption among alcohol consumers depending on survey design, health, and birth cohort

		Bivariate		Within categories		Model 4		Model 5
		OR	p-value	OR	p-value	OR	p-value	OR	p-value
Survey design				Model 1
Interview succession (linear)		0.96	0.079	0.98	0.353	0.97	0.289	0.99	0.585
Proxy		0.44	<0.001	0.49	0.001	0.77	0.293	0.80	0.359
Telephone		0.62	0.004	0.85	0.419	0.84	0.397	0.72	0.106
Health				Model 2
Living in an institution (Women)		0.29	<0.001	0.39	0.002	0.46	0.019	0.45	0.020
Living in an institution (Men)		0.49	0.036	0.68	0.273	0.80	0.544	0.74	0.423
ADL limitation		0.53	0.000	0.77	0.127	0.81	0.219	0.86	0.400
Mobility problem		0.66	0.001	0.78	0.056	0.80	0.077	0.77	0.038
Birth cohort				Model 3
Period	1992	1.00	ref	1.00	ref			1.00	ref
	2002	1.15	0.395	0.89	0.620			1.18	0.481
	2011	1.88	<0.001	1.16	0.625			1.89	0.044
Birth cohort (linear)		1.40	<0.001	1.32	0.036			1.04	0.801
Age (linear)		0.79	0.084	^b

Significant estimates (p<0.05) are in bold ^aFrom ordered logistic regression models, which provide the average increase of the odds ratio for reporting one higher category, e.g. for weekly rather than monthly. The outcome had three levels (Seldom, monthly or weekly). The assumption of equal effect sizes (also called proportional odds/parallel lines) was tested with partial proportional odds models (Stata command gologit2). The assumption was not violated for any of the independent variables (p > =0.084 in model 5)
^bAs age, cohort and period cannot be analysed in the same model, and because the period change was not related to changed age distribution over the years, age was excluded in the full model

ISSN: 1471-2458