Respondents with missing information on any variable described above are excluded.

Logistic regression in Stata 12 SE is used, and coefficients are average marginal effects (AME) predicted with the margins option. Contrary BKM120 to what is often believed, log-odds ratios or odds ratios are not comparable across studies or models ( Mood, 2010 and Wooldridge, 2002). Therefore, AME are reported, which are easily interpretable as the average impact on the probability (0–1) of good health. For categorical variables, AME give the discrete difference in the probability of good health between the relevant category and the reference group. As the outcome is restricted to be 0 or 1 the estimated effects are not additive: If a person has many risk factors, the measured outcome can still not be worse than “not good.” FGFR inhibitor The predicted probabilities of

good health in 2000 at different combinations of risk factors will therefore also be shown, using a type case, and varying the statistically significant lifestyle factors one by one and in combination for this case. The type case is a woman of average age, income and education, who usually drinks less than two glasses, eats vegetables daily, is not overweight, and does not see friends and family often (smoking, exercise and social support are set to vary). Because of Modulators sample size restrictions, response categories for some variables have been collapsed. In these cases, different categorizations have been tested, and those reported give the most robust results. Descriptives for all variables are given in Table 1. Recall that all respondents had good health in 1991, so the 20% reporting less than good self-rated health in 2000 or 2010 have seen deterioration. There are equal shares of men and women, and the average age in 1991 is 38 for respondents observed in 2000 and 36 for those observed in 2010 (this decline is explained by panel ageing, as those who remained in 2010 were younger in 1991 than those who remained in 2000). Around 30% are single households, and 28% are overweight in 1991. A majority, 74%, exercise each week, and around 60%

eat vegetables every day. 49% have never smoked, and around 30% currently smoke. Less than 10% never STK38 drink alcohol, and of those who drink, around half usually drink more than a couple of glasses. Around half the sample see friends often and an equal share see family often. Only 4% lack social support. Table 2 gives regression results for self-rated health in 2000 (models 1A–1B) and in 2010 (models 2A–2B). In both cases, model A includes lifestyle variables, and model B additionally includes control variables. Model 1A shows that weekly exercise, usually drinking more than two drinks, and seeing friends often in 1991 are positively related to health in 2000 (statistically significant, P < 0.05), while smoking and lack of social support are negatively related to health (P < 0.05).