Marcel Slockers

32 Chapter 2 linear predictor consists of two categorical variables operationalized by dummy x1 for being female (male is the baseline) and several dummies for age group: x2 being in the second age group (youngest age group is baseline), x3 etc. Exp(alpha) is the rate for males in the youngest (baseline) age group. The effect of each variable is corrected for the other variables. We did not correct for period, as mortality did not differ significantly within the 10-year of follow up. Poisson regression yields rate ratios (RRs) which in our model indicate the increase/decrease in the mortality rate relative to the reference category (e.g. homeless women as compared to homeless men). Age was grouped in 10-year age groups. Poisson regression with follow-up time by calendar year and age yields the same results as Cox regression with age as baseline hazard if the variable age is truncated to 10-year age groups. 24-26 Poisson regression including age*gender interactions yields the same results as an age-sex stratified analyses, but also provides overall p-values for the significance of the interaction. Next, we compared mortality between homeless people and the general Rotterdam population. We first constructed a similar dataset for homeless people and the general Rotterdam population with person-years and mortality by 5-year age group and sex for the period 2001–2010. Therefore, we aggregated person-years at risk and mortality among the homeless population across ages into 5-year age groups. Data on the general Rotterdam population were obtained from Statline and had the same format. Poisson regression with person-years as rate multiplier was used to assess the association between homelessness and mortality, and to assess whether this association differed by age and sex. This yields RRs comparing mortality rates among homeless men with those among men in the general population, corrected for age. The same model yields age-spe- cific RRs, corrected for gender. These RRs can be interpreted as standardized mortality ratios, 27 and indicate the amount of excess mortality in the homeless population relative to the general population. We calculated overall RRs corrected for age (5-year age groups) and sex, gender-specific RRs corrected for age, and age-specific RRs corrected for gender. To increase the power we defined these interactions between homelessness and age using broader age groups (<45, 45–59, 60+ years). Finally, we used the actuarial method to calculate remaining life expectancy by sex for homeless men and women, and for men and women in the general population for different starting ages. Life expectancies are not affected by the population structure and therefore can be directly compared between populations or population groups. Confidence estimates were obtained using the method proposed by Chiang. 28 To further examine the difference in life expectancy between homeless people and the general population, we assessed the contribution of different age groups to the disparity in remaining life expectancy using the