Smoking and cognitive functioning at older ages: Evidence from the Health and Retirement Study

Smoking and cognitive functioning at older ages: Evidence from the Health and Retirement Study

The Journal of the Economics of Ageing xxx (2015) xxx–xxx Contents lists available at ScienceDirect The Journal of the Economics of Ageing journal h...

458KB Sizes 2 Downloads 89 Views

The Journal of the Economics of Ageing xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

The Journal of the Economics of Ageing journal homepage: www.elsevier.com/locate/jeoa

Full Length Article

Smoking and cognitive functioning at older ages: Evidence from the Health and Retirement Study Padmaja Ayyagari a,⇑, Asia Sikora Kessler b a b

Department of Health Management and Policy, University of Iowa, 145 N. Riverside Drive, Iowa City, IA 52242, USA Department of Health Promotion, Social and Behavioral Health, University of Nebraska Medical Center, 984365 Nebraska Medical Center, Omaha, NE 68198-4365, USA

a r t i c l e

i n f o

Article history: Available online xxxx Keywords: Smoking Cognitive functioning Older adults Fixed effects models

a b s t r a c t It has long been established that smoking has detrimental health effects on the human body and leads to early mortality. However, the evidence regarding the link between smoking behavior and cognitive function in old age is mixed. While nicotine has been shown to improve cognitive functioning in some medical studies, current smokers typically perform worse than non-smokers on cognitive tests in population based studies. We use the Health and Retirement Study to evaluate the role of unobserved factors in explaining these conflicting findings. Results from individual fixed effects models show that much of the lower cognitive function of current and former smokers relative to never smokers can be attributed to unobserved differences between these groups. Ó 2015 Elsevier B.V. All rights reserved.

Introduction More than 5 million individuals currently have Alzheimer’s Disease (AD) and this number is projected to increase substantially over the next few years as the U.S. population ages (Alzheimer’s Association Report, 2012). While age and the APOE e4 genotype (Farrer et al., 1997) have been identified as important risk factors for dementia, much remains unknown about the causes and mechanisms leading to cognitive decline in old age. Recent studies suggest that a complex interaction of genetics, environment and lifestyle likely cause AD and related dementia. In this study, we investigate the role of smoking behavior in explaining cognitive functioning among older adults. We focus our analysis on cognitive functioning, rather than a clinical diagnosis of AD, since mild cognitive impairments are often a precursor to more severe declines in cognitive functioning or the onset of dementia. Previous studies have examined the relationship between smoking and the risk of dementia with mixed results. While some studies have found that smokers have a lower risk of Alzheimer’s disease (Fratiglioni and Wang, 2000) or that there is no association between smoking and dementia (Yamada et al., 2003), others have found that smoking increases the risk of dementia (Almeida et al., 2002; Tyas et al., 2003; Whitmer et al., 2005). Peters et al. (2008) conducted a systematic review of the literature on smoking, ⇑ Corresponding author. E-mail addresses: [email protected] (P. Ayyagari), asia.sikora@ unmc.edu (A. Sikora Kessler).

dementia and cognitive decline in the elderly, and concluded that current smoking increases the risk of Alzheimer’s disease (AD) but not of other dementias nor of cognitive decline. This result was reinforced by another review article by Qiu et al. (2009), who suggest that survival bias might explain any protective effect identified by earlier studies. Anstey et al. (2007) found that current smoking status was associated with a decline in cognitive function based on a meta-analysis of 19 studies. In contrast, nicotine has been shown to have a protective effect on AD (Ulrich et al., 1997). One of the key features of AD is an increased production of b-amyloid plaques in the brain (Alzheimer’s Disease Education & Referral (ADEAR) Center, 2011), and a growing medical literature finds evidence that nicotine inhibits the formation of such plaques (Salomon et al., 1996; Zanardi et al., 2002; Srivareerat et al., 2011). Studies that have delivered nicotine through skin patches have shown improved cognitive functioning and support similar evidence from animal studies (Wilson et al., 1995; Rezvani and Levin, 2001; Newhouse et al., 2012). On the other hand, Almeida et al. (2011) compared smokers enrolled in a smoking cessation trial with never smokers and found that smokers who quit exhibited cognitive declines similar to never smokers, while ongoing smokers exhibited greater cognitive declines. Given these conflicting results, there is a need for further research to better understand the relationship between smoking behaviors and cognition. In addition to the plaque inhibiting properties of nicotine, there may be other pathways through which smoking is associated with cognition in old age. One such pathway is through the impact of smoking on cardiovascular disease. Smoking is known to be

http://dx.doi.org/10.1016/j.jeoa.2015.06.001 2212-828X/Ó 2015 Elsevier B.V. All rights reserved.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

2

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx

correlated with increased cardiovascular risk which may in turn lead to an increased risk of vascular dementia. Secondly, there may be ‘‘third’’ factors that are associated with both smoking behavior and cognitive ability leading to a spurious correlation between the two. For example, education has been shown to be strongly correlated with cognition in old age (Stern et al., 1994; Roe et al., 2007; Banks and Mazzonna, 2012) and at the same time individuals with high educational attainment are also less likely to be smokers (Kenkel, 1991; de Walque, 2007, 2010). Therefore, accounting for such confounding factors will be important in any study assessing the relationship between smoking and cognitive functioning. In this study, we take advantage of the longitudinal nature of the HRS to estimate intra-individual or within-person effects of smoking on cognition. Specifically, we estimate an individual fixed effects (FE) regression model, which allows us to account for any unobserved time-invariant factors that may be correlated with both smoking and cognition. Our study contributes to the existing literature along several dimensions. First, we use a nationally representative sample of older adults in contrast to most previous studies that have often used small, selective samples. Second, most previous research has used a study design that compares the cognition of smokers with non-smokers. As mentioned above, the main concern with such analyses is that smokers may differ from non-smokers in unobserved ways that also impact their cognitive functioning. We extend this literature by using panel data methods to evaluate the effect of within person changes in smoking behavior on changes in cognition and compare our results to results from a cross-sectional study design. Third, we use a rich set of information on socioeconomic factors and health status to evaluate their role in explaining associations between smoking and cognition. Finally, we assess the extent to which our results may be driven by differential rates of survival between smokers and non-smokers or by proxy respondent status. We find that much of the difference in cognitive functioning between current smokers and never smokers can be attributed to unobserved factors. In addition, we find that current smokers perform better than former smokers on some measures of cognition once unobserved factors are taken into account. Our results suggest that there is a need for further research on the role of nicotine in influencing cognition and the onset of dementia.

Data We use data from the 1998 through 2008 waves of the Health and Retirement Study (HRS). The HRS is a nationally representative, longitudinal survey of individuals over 50 years and their spouses. The HRS initially sampled persons in birth cohorts 1931 through 1941 and conducted follow up interviews biennially. In 1998, persons from the 1924 to 1930 cohort and the 1942 to 1947 cohort were added to the original sample. In 2004, persons from the 1948 to 1953 cohort were added to the survey. We use data from the RAND HRS (version L) data, which is a longitudinal data file that includes cleaned versions of the most frequently used HRS variables. Although survey data is available beginning in 1992, the measures of cognition changed over survey waves and have been consistently measured since 1998. Therefore, we do not use data from the prior waves of the survey. The RAND HRS dataset is a sample of 30,671 individuals who were interviewed at various waves of the HRS. To obtain a relatively homogenous sample, the main analysis is restricted to individuals who are between 50 and 100 years of age, reducing the analysis sample to 25,982 persons. Individuals not residing in the U.S. at the time of the interview are omitted from the analysis as are observations with missing values for key

analysis variables such as cognition, smoking status and demographics. This leaves us with a sample of 24,164 persons. Since our estimation method exploits the panel nature of the HRS data, we further restrict the sample to persons who provide at least two waves of data. The final analysis sample includes 21,226 individuals and 97,307 person-wave observations. Table 1 presents summary statistics for the analysis sample. Measures of cognitive functioning The HRS obtained detailed measures of cognition in terms of learning and memory, which are considered the earliest and core signs for dementia (Herzog and Wallace, 1997). The HRS cognitive tests were designed to capture various dimensions of cognition: knowledge, reasoning, orientation, calculation, and language (Ofstedal et al., 2005). The RAND HRS data set provides summary cognitive scores based on survey questions, with imputed values for missing observations (Fisher et al., 2012). We use six measures of cognitive functioning scores obtained from this dataset. Table 1 presents summary statistics and Fig. 1 presents the histograms for each of the six measures. The measures are defined as follows: Self memory This measure of memory is based on a survey question, in which respondents are asked to rate their memory at the present time. Answers are recoded so that higher values represented better memory, and range from 1 (representing poor memory) to 5 (representing excellent memory). Close to 30% of the sample report very good to excellent memory with the majority of individuals reporting good memory. Past memory The rate of decline in cognitive functioning may be a more important indicator of dementia than the level of cognitive functioning (Jorm and Jacomb, 1989) and therefore respondents are asked how their current memory compares to their memory two years back or at the time of their previous interview. Answers were re-coded as 1 for ‘‘worse now’’, 2 for ‘‘about the same’’ and 3 for ‘‘better now’’. On average, 21% of the sample reports a worsening of memory relative to the previous interview while the majority (77%) reports no change. Serial 7 This variable is an objective measure of the individual’s working memory and is based on a task in which respondents are asked to subtract 7 from 100 and to continue subtracting 7 from each subsequent number for a total of five times. The score is the count of correct subtractions across the five trials, with each subtraction being assessed independently. The score ranges from 0 to 5. Over half the sample makes at least one mistake while performing this task and about 8% of individuals do not get a single correct answer. Total word recall This measure of episodic memory is based on two tasks – immediate word recall and delayed word recall. Interviewers read a list of 10 nouns to the respondent who is then asked to recall as many words as possible in any order. After approximately 5 min, during which time respondents are asked other survey questions, they are asked to repeat the task. The score is the count of the number of words that were correctly recalled both times and ranges from 0 to 20. This measure has a mean score close to 10. Mental status This score is a combination of various tests that are designed to measure knowledge, language and orientation. It is the sum of the

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

3

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx Table 1 Summary statistics. Full sample Mean or percent (Std. Dev.)

Never smokers Mean or percent (Std. Dev.)

Former smokers Mean or percent (Std. Dev.)

Current smokers Mean or percent (Std. Dev.)

Measures of cognition Self-memory = excellent Self-memory = very good Self-memory = good Self-memory = fair Self-memory = poor Past memory = better Past memory = same Past memory = worse Serial 7 Total recall Mental status Total cognition

5.87 23.70 43.06 22.64 4.73 2.40 76.72 20.87 3.51 (1.69) 9.87 (3.64) 12.67 (2.54) 21.79 (5.27)

5.73 24.81 44.16 21.25 4.05 2.37 77.70 19.94 3.47 (1.72) 9.98 (3.68) 12.57 (2.62) 21.77 (5.42)

5.60 22.92 42.80 23.59 5.10 2.25 75.53 22.22 3.60 (1.65) 9.72 (3.62) 12.79 (2.47) 21.80 (5.17)

7.12 22.80 40.63 23.82 5.63 2.98 77.48 19.54 3.36 (1.74) 10.03 (3.54) 12.54 (2.49) 21.88 (5.14)

Measures of smoking Never smoker Former smoker Current smoker Number of cigarettes per day

42.15 43.53 14.33 N/A

N/A N/A N/A N/A

N/A N/A N/A N/A

N/A N/A N/A 15.77 (11.67)

Other covariates Age Male Black Other race Hispanic Years of education Married or partnered N Persons

67.66 (10.03) 40.01 13.69 3.97 7.93 12.37 (3.19) 64.53 97,307 21,226

68.20 (10.53) 27.25 13.73 4.32 8.87 12.45 (3.28) 63.99 41,013 8878

68.58 (9.71) 51.51 12.71 3.46 7.21 12.47 (3.17) 67.40 42,353 10,146

63.28 (8.16) 42.61 16.56 4.50 7.34 11.82 (2.97) 57.38 13,941 4178

F G VG Self Memory

Density .2 .3 .1

E

Worse Same Better Past Memory

2 4 Serial 7

6

Density .04 .06

.08

0

0

0

0

.02

.1

.05

Density

Density

.2

.1

.3

.15

P

0

0

0

.1

.2

Density .4

Density .2 .3

.6

.4

.4

.8

Notes: The following variables are not available for all observations due to missing values: mental status, total cognition, former smoker, current smoker and number of cigarettes per day.

0

5

10 15 Total Recall

20

0

5 10 Mental Status

15

0

10 20 30 Total Cognition

40

Fig. 1. Histograms of measures of cognitive functioning.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

4

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx

scores for the serial 7, backwards counting from 20, object naming, date naming and president/vice-president naming. Respondents are asked to count backwards as quickly as possible beginning with the number 20. Answers were coded as 0 for incorrect, 1 for correct on the second try and 2 for correct on the first try. In addition, respondents are asked to report the month, day, year and day of the week during their interview. They are also asked to name the object they would ‘‘usually use to cut paper’’ and ‘‘the kind of prickly plant that grows in the desert’’. Each of these answers was coded as 1 for a correct response and 0 for an incorrect one. Individuals are also asked to name the current President and Vice President of the United States and answers were coded as 1 for getting each last name right and 0 otherwise. The mental status score ranges from 0 to 15, with a higher value representing better cognition. The questions on the President and Vice President of the United States were only asked of new-interviewees and re-interviewees who are 65 years or older. Therefore, the mental status and total cognition scores are only available for a subsample (n = 61,582 or 63.29% of the full sample). Approximately 70% of individuals respond incorrectly to at least one of the questions used to create this measure, with most persons getting one to two questions wrong. Total cognition This score sums the mental status score and the total word recall scores and ranges from 0 to 35, with a higher value representing better cognition. The mean score on this measure is 21.8. Measures of smoking behavior All survey respondents are asked whether they had ever smoked cigarettes. Approximately 58% of the sample report that they have smoked cigarettes once in their lifetime (‘‘ever smokers’’). These individuals are then asked if they smoke cigarettes now and approximately 25% of ever smokers report that they do (‘‘current smokers’’). In part of the analysis described below we restrict the sample to ever smokers and examine the impact of current smoking status on cognition. Current smokers were also asked to report the quantity of cigarettes smoked per day, which could be reported in number of cigarettes, packs or cartons. We calculate the number of cigarettes smoked per day by assuming that each pack contains 20 cigarettes and each carton contains 200 cigarettes. Conditional on being a current smoker, the average number of cigarettes smoked per day is 15.8. Other variables All regressions include a basic set of covariates: age, age squared, years of education, and binary indicators for male, black and other race (white being the reference category), Hispanic ethnicity and being married or partnered (with all other marital statuses such as separated, divorced, widowed or never married forming the reference group). Regressions also include indicators for census region of residence. In addition, models based on objective measures of cognitive functioning include a measure of prior exposure to capture learning effects. Rodgers et al. (2003) show that persons with prior exposure to the cognitive tests perform better on these tests at subsequent interviews. Therefore, better performance due to learning about the tests is likely to be confounded with true changes in cognitive functioning. To account for such effects, we include a measure which is a running sum of the number of times each respondent has answered the cognitive questions. Since new cohorts were added to the HRS at different times, the prior exposure variable is not perfectly correlated with survey wave or year effects. In addition to these basic set of covariates, some regressions include additional measures such as total

household income (inflation adjusted) and retirement status. We also include various measures of health such as binary indicators for self-rated health, Center for Epidemiologic Studies Depression (CESD) score, count of activities of daily living (ADL) limitations and count of instrumental activities of daily living (IADL) limitations. In addition, we include indicators for ever having been diagnosed with high blood pressure, diabetes, heart disease, stroke or psychological problems. Methods We begin by examining the association between smoking and cognitive ability using ordinary least squares (OLS) regression. The regression equation for person i at time t is given by:

Cognitionit ¼ a0 þ a1 Current smokerit þ a2 Former smokerit þ a3 X it þ a4 Yeart þ eit

ð1Þ

The key explanatory variables are binary indicators for current and former smoking status (with never smoker as the reference category). Regressions include year fixed effects (denoted by Yeart ) that account for a secular trend in cognition as well as the basic set of socio-demographic variables described above. To account for geographic variation in environmental factors that may affect both smoking and cognition, we also include binary indicators for census region of residence. These covariates are represented by the vector X. However, there may be other omitted variables that are confounded with smoking decisions, so that a1 and a2 cannot be interpreted as causal effects. To account for potential confounders, we estimate an individual fixed effects (FE) model specified as follows:

Cognitionit ¼ b0 þ b1 Current smokerit þ b2 Former smokerit þ b3 X it þ b4 Yeart þ li þ #it

ð2Þ

The fixed effects model accounts for unobserved factors that are time-invariant or person specific (represented by li ), and that may be correlated with the idiosyncratic error term (represented by #it ). This method essentially estimates the impact of within person changes in smoking status on within person changes in cognitive functioning. To test whether the FE model is indeed the appropriate model for our application, we use the Hausman specification test to compare a random effects model with the FE model.1 One concern with the above analysis is that never smokers are likely to be substantially different from former or current smokers in unobserved ways. Therefore, we also estimate OLS and FE models for the sample of ever smokers. The key explanatory variable in these analyses is a binary indicator for current smoking status (with former smokers being the reference category). While the FE models are an improvement over existing studies, the obtained estimates may still not represent the causal effect of smoking behaviors on cognition due to four potential sources of bias. One is the presence of time-varying unobserved factors that are correlated with both smoking behaviors and cognition. For example, heart disease and stroke are well known to be associated with smoking and several recent studies have shown that they are also risk factors for dementia (Savva and Stephan, 2010; Hjelm et al., 2012). Other factors such as retirement have been shown to affect cognitive functioning (Rohwedder and Willis, 2010; Bonsang et al., 2012; Mazzonna and Peracchi, 2012) and retirement has also been linked to smoking behaviors (Henkens et al., 2008). If, 1 The random effects (RE) model accounts for unobserved heterogeneity but assumes that the unobserved effects are not confounded with smoking status or other observed covariates. The Hausman test is based on the principle that in the presence of unobserved confounders, the RE model produces biased and inconsistent estimates while the FE estimates are consistent.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx

as in the case of stroke, these unobserved factors are positively associated with smoking and negatively with cognition, then our estimated parameters will be biased downwards. Although our estimation methods cannot conclusively rule out these biases, we perform robustness checks to evaluate the sensitivity of our estimates to such biases. Taking advantage of the rich set of information available in the HRS, we estimate regressions that include a wide range of health measures and other variables that have been shown to be associated with cognition. The second source of bias is survivorship bias. Survivorship bias results from the mortality differential between smokers and non-smokers. Life expectancy is 10 years shorter for smokers relative to never smokers (Jha et al., 2013), but smoking cessation confers substantial benefits, adding 6.1 to 8.5 years to life expectancy for those who quit at age 35 (Taylor et al., 2002). The lower life expectancy for smokers raises the concern that non-smokers are over-represented in our sample of older adults and heavy smokers are under-represented. If smoking decreases cognition then our estimates will be biased upwards. To address survivorship bias, we examine the impact of smoking behaviors separately for persons aged 50 to 65 and those aged 66 to 85 years. Given an average life expectancy of 78.7 years in the U.S. (Murphy et al., 2013), we expect that the 66–85 age group will exhibit a larger mortality differential between smokers and non-smokers compared to the 50– 65 age group. Therefore, any difference in the estimates obtained from these subsamples would indicate the magnitude of the survivorship bias. The third source of bias is the exclusion of individuals represented by a proxy from our sample. The questions on cognitive function are only asked to self-respondents in the HRS. If cognitive decline is associated with the switch from self-respondent to proxy-respondent status, our estimates would be biased toward better cognitive function if smoking reduces cognition. To assess the potential role of this bias, we examine whether smokers are more likely to be represented by a proxy in each year of our sample. The final source of bias is reverse causality. If, for example, individuals with better cognition are less likely to be smokers or more likely to quit early then the estimate obtained from Eq. (2) would not represent the causal effect. Unfortunately, publicly available HRS data does not contain information needed to address this bias or to assess its magnitude. Despite this limitation, our paper makes a significant contribution to the literature by addressing individual specific unobserved variables. In the concluding section, we discuss additional methodological and conceptual issues that future research should address.

Results Table 2 presents the results from OLS regressions for the full sample. Across all six measures of cognition, current smokers perform worse than never smokers and the total cognition score is 0.24 units lower for current smokers relative to never smokers. We also find evidence that former smokers have worse cognition than never smokers, as measured by self-reports and the serial 7 task which measures working memory. However, there are no significant differences between former smokers and never smokers in memory (as measured by the total recall score) and in knowledge, language or orientation (as measured by the mental status score). The strong advantage of never smokers relative to current smokers and the weaker evidence of advantage relative to former smokers, suggest that smoking cessation may be successful in restoring cognitive health. At the same time, it is highly probable that these effects mask unobserved differences between smokers and non-smokers. As mentioned above, many of the

5

factors that predict smoking status such as socioeconomic status, mental health and chronic health conditions have also been shown to be associated with cognitive functioning. Consistent with the literature showing a positive link between education and cognition (Banks and Mazzonna, 2012), we find that individuals with fewer years of education have worse cognition across all six measures. Blacks and Hispanics self-report lower memory and perform worse on all the objective measures, however, Blacks self-report better cognition relative to the past. The impact of age on cognition is a bit complex. In analysis not shown, we estimated a model using only a linear age term. Consistent with prior literature, we find that on average total cognition declines with age. The coefficient on age was negative and significant at the 1% level. In the results presented below, we include both age and age squared in the regressions to capture nonlinear impacts of age. Both age and age squared are statistically significant in the results presented below. Thus, while on average cognition declines with age, the impact of age appears to be nonlinear. Further analysis using categorical age variables (50– 59 as the reference, and 60–69, 70–79, 80–89, 90 or older) shows that declines in cognition are much larger after age 80 compared to declines in cognition at younger ages. In Table 3, we present results from the FE models which account for person specific unobserved factors. For all six measures of cognition, the Hausman test rejects the null hypothesis that the RE model is the appropriate one (not shown). This suggests that unobserved confounders play a significant role in explaining the association between smoking and cognition observed in the OLS models. Therefore, the FE model is our preferred specification. The effect of smoking status on cognition is no longer statistically significant for any of the measures except for mental status, and in fact has a positive sign for past memory, serial 7 and word recall. In the case of mental status, scores are 2.5 units lower for former and current smokers compared to never smokers. Overall, these results suggest that while smokers have worse cognition than non-smokers on average, most of this difference is driven by person-specific unobserved factors. For example, factors such as childhood environment or nutrition have long term impacts on cognitive health in old age (Case and Paxson, 2009; Maurer, 2010; Guven and Lee, 2013) and are also associated with health behaviors such as smoking (Lacey et al., 2011). Given that non-smokers may differ substantially from smokers in unobserved ways, we focus the remaining analyses on the sample of ever smokers. Some of the unobserved factors that determine smoking initiation are likely to be common for former and current smokers, so that dropping never smokers reduces the potential for omitted variable bias. However, there could be other unobserved factors that cause some persons to quit smoking and others to continue smoking even at older ages. Results from OLS regressions are presented in Table 4 and results from FE models are presented in Table 5. The findings from the OLS regressions are somewhat mixed. Relative to former smokers, current smokers obtain significantly lower scores on all four of the objective measures of cognition but they self-report better memory compared to the past. However, when we account for unobserved individual specific factors we find that current smokers have significantly better self-reported memory and they perform better on the word recall tasks compared to former smokers. Current smokers also have better scores on the serial 7, mental status and total cognition measures but these effects are not statistically significant at conventional levels. The estimates are consistent with findings from the medical literature showing that nicotine improves attention and other aspects of cognitive functioning. Our results also suggest that the lower cognitive functioning of current smokers identified in previous epidemiologic literature is likely biased due to unmeasured confounders that also affect brain health.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

6

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx

Table 2 Impact of smoking status on cognition: OLS models.

Current smoker Former smoker Age Age squared Male Black Other race Hispanic Years of education Married/partnered

(1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

0.0447*** (0.0153) 0.0583*** (0.0110) 0.0448*** (0.0055) 0.0003*** (0.0000) 0.0273** (0.0109) 0.1733*** (0.0159) 0.0199 (0.0268) 0.1142*** (0.0219) 0.0639*** (0.0018) 0.0311*** (0.0108)

0.0117* (0.0063) 0.0275*** (0.0047) 0.0007 (0.0023) 0.0000 (0.0000) 0.0233*** (0.0045) 0.0211*** (0.0066) 0.0188 (0.0115) 0.0055 (0.0092) 0.0025*** (0.0008) 0.0204*** (0.0046)

4.2583*** (0.1897) 97,307 21,226

1.9524*** (0.0817) 97,307 21,226

0.0777*** (0.0252) 0.0355** (0.0179) 0.0493*** (0.0091) 0.0005*** (0.0001) 0.2988*** (0.0175) 1.0395*** (0.0285) 0.4285*** (0.0497) 0.2500*** (0.0412) 0.1822*** (0.0030) 0.1039*** (0.0180) 0.0366*** (0.0072) 0.1035 (0.3138) 97,307 21,226

0.2034*** (0.0469) 0.0372 (0.0346) 0.2085*** (0.0174) 0.0025*** (0.0001) 1.1326*** (0.0338) 1.2412*** (0.0480) 0.7281*** (0.0843) 0.2172*** (0.0683) 0.3237*** (0.0055) 0.1321*** (0.0344) 0.1969*** (0.0135) 4.1022*** (0.6078) 97,307 21,226

0.0753* (0.0441) 0.0427 (0.0303) 0.2462*** (0.0145) 0.0021*** (0.0001) 0.3073*** (0.0295) 1.7458*** (0.0517) 0.7200*** (0.0946) 0.4841*** (0.0735) 0.2878*** (0.0055) 0.1322*** (0.0298) 0.0905*** (0.0128) 2.4627*** (0.5053) 61,582 19,126

0.2443*** (0.0878) 0.0021 (0.0610) 0.5113*** (0.0274) 0.0050*** (0.0002) 0.8189*** (0.0594) 3.0303*** (0.0940) 1.4548*** (0.1707) 0.7224*** (0.1333) 0.5942*** (0.0102) 0.2223*** (0.0597) 0.2056*** (0.0251) 4.9330*** (0.9646) 61,582 19,126

Prior exposure Constant N Persons

Regressions also include year and census region of residence dummies. Robust standard errors presented in parentheses are clustered at the individual level. * p < 0.10. ** p < 0.05. *** p < 0.01.

Table 3 Impact of smoking status on cognition: FE models.

Current smoker Former smoker Age Age squared Married/partnered

(1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

0.3407 (0.3236) 0.4417 (0.3239) 0.0193** (0.0096) 0.0001** (0.0000) 0.0206 (0.0131)

0.1260 (0.0819) 0.0822 (0.0823) 0.0013 (0.0056) 0.0001*** (0.0000) 0.0063 (0.0076)

2.5949*** (0.5411) 97,307 21,226

2.3260*** (0.3017) 97,307 21,226

0.0462 (0.2731) 0.0285 (0.2732) 0.1280*** (0.0153) 0.0011*** (0.0001) 0.0114 (0.0204) 0.0424* (0.0217) 0.0673 (0.8339) 97,307 21,226

0.8959 (0.7716) 0.6789 (0.7740) 0.5204*** (0.0347) 0.0040*** (0.0002) 0.0389 (0.0456) 0.0615 (0.0508) 6.6542*** (1.9174) 97,307 21,226

2.5067*** (0.8694) 2.5306*** (0.8722) 0.8403*** (0.0438) 0.0058*** (0.0002) 0.0051 (0.0413) 0.2223*** (0.0303) 15.3339*** (2.1975) 61,582 19,126

2.0630 (1.6110) 2.3223 (1.6154) 1.4060*** (0.0834) 0.0099*** (0.0004) 0.0791 (0.0797) 0.2191*** (0.0577) 24.6481*** (4.3071) 61,582 19,126

Prior exposure Constant N Persons

Regressions also include year and census region of residence dummies. Robust standard errors presented in parentheses are clustered at the individual level. * p < 0.10. ** p < 0.05. *** p < 0.01.

While the FE models are able to address biases due to time-invariant unobserved factors, they do not account for time-varying unobserved variables affecting both smoking and cognition. Particularly in the case of older individuals, events such as health shocks or decreased labor force participation are common and such events have been shown to affect both smoking and cognition (Falba, 2005; Khwaja et al., 2006; Keenan, 2009; Savva and Stephan, 2010). To account for such time varying factors, we re-estimate the FE models adding extensive controls for health, income and retirement status (see Table 6). The positive effect of current smoking on cognition remains even after adjusting for

these variables. As expected, worse health has a significant negative effect on cognitive functioning. Poorer self-rated health and ADL and IADL limitations are associated with worse scores for all six measures of cognitive functioning. A history of stroke reduces the total cognition score by 0.9 points. Psychological problems and a higher number of depressive symptoms (as measured by the CESD score) are associated with worse cognition, but not for all measures. Consistent with the findings of Rohwedder and Willis (2010), we find that retirement is associated with worse cognition. Log income has a positive effect. Overall, the FE results suggest that conditional on having been a smoker, current smoking

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

7

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx Table 4 Impact of current smoking among ever smokers on cognition: OLS models.

Current smoker Age Age squared Male Black Other race Hispanic Years of education Married/partnered

(1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

0.0195 (0.0150) 0.0445*** (0.0078) 0.0003*** (0.0001) 0.0220 (0.0138) 0.1551*** (0.0213) 0.0215 (0.0376) 0.0413 (0.0295) 0.0680*** (0.0025) 0.0216 (0.0143)

0.0171*** (0.0062) 0.0022 (0.0034) 0.0000 (0.0000) 0.0320*** (0.0058) 0.0239*** (0.0086) 0.0240 (0.0161) 0.0084 (0.0125) 0.0033*** (0.0010) 0.0172*** (0.0061)

4.1223*** (0.2699) 56,294 12,360

1.9700*** (0.1165) 56,294 12,360

0.1098*** (0.0241) 0.0404*** (0.0124) 0.0004*** (0.0001) 0.3247*** (0.0223) 1.0430*** (0.0371) 0.4943*** (0.0670) 0.1701*** (0.0552) 0.1848*** (0.0039) 0.1054*** (0.0233) 0.0338*** (0.0094) 0.4406 (0.4261) 56,294 12,360

0.1655*** (0.0456) 0.1765*** (0.0240) 0.0023*** (0.0002) 1.1602*** (0.0431) 1.2242*** (0.0626) 0.6237*** (0.1150) 0.1122 (0.0927) 0.3288*** (0.0072) 0.1427*** (0.0448) 0.2029*** (0.0180) 5.1024*** (0.8330) 56,294 12,360

0.1172*** (0.0427) 0.2245*** (0.0197) 0.0019*** (0.0001) 0.3225*** (0.0374) 1.7045*** (0.0659) 0.6547*** (0.1273) 0.4144*** (0.0980) 0.2851*** (0.0071) 0.1572*** (0.0385) 0.0983*** (0.0167) 3.3055*** (0.6805) 35,107 11,111

0.2512*** (0.0856) 0.4657*** (0.0373) 0.0047*** (0.0003) 0.8480*** (0.0757) 2.9695*** (0.1219) 1.3251*** (0.2341) 0.5346*** (0.1799) 0.5981*** (0.0133) 0.2575*** (0.0778) 0.2128*** (0.0324) 6.4914*** (1.3055) 35,107 11,111

Prior exposure Constant N Persons

Regressions also include year and census region of residence dummies. Robust standard errors presented in parentheses are clustered at the individual level. *** p < 0.01.

Table 5 Impact of current smoking among ever smokers on cognition: FE models.

Current smoker Age Age squared Married/partnered

(1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

0.0956*** (0.0170) 0.0248* (0.0131) 0.0001** (0.0001) 0.0277 (0.0176)

0.0411*** (0.0092) 0.0088 (0.0075) 0.0001*** (0.0000) 0.0055 (0.0101)

2.0627*** (0.6879) 56,294 12,360

1.8180*** (0.3910) 56,294 12,360

0.0102 (0.0253) 0.1357*** (0.0204) 0.0012*** (0.0001) 0.0514* (0.0267) 0.0691** (0.0272) 0.3490 (1.0764) 56,294 12,360

0.2087*** (0.0575) 0.5172*** (0.0459) 0.0038*** (0.0002) 0.0134 (0.0591) 0.1437** (0.0642) 6.9829*** (2.4402) 56,294 12,360

0.0147 (0.0601) 0.8445*** (0.0587) 0.0057*** (0.0003) 0.0840 (0.0543) 0.1437** (0.0642) 17.4617*** (2.8118) 35,107 11,111

0.2357* (0.1212) 1.3991*** (0.1120) 0.0095*** (0.0006) 0.0416 (0.1034) 0.2750*** (0.0757) 27.5475*** (5.5052) 35,107 11,111

Prior exposure Constant N Persons

Regressions also include year and census region of residence dummies. Robust standard errors presented in parentheses are clustered at the individual level. * p < 0.10. ** p < 0.05. *** p < 0.01.

status is associated with better cognitive functioning for some measures of cognition. These findings are consistent with the hypotheses that nicotine improves cognitive functioning and inhibits the formation of senile plaques. Next, we examine whether smoking intensity affects cognition for the sample of current smokers. Results from FE models estimating the impact of the average number of cigarettes smoked per day on cognition are presented in Table 7. We do not find an effect of smoking intensity for any of the measures of cognition, except for word recall. In the case of word recall, increased smoking intensity is associated with better recall but this benefit appears to diminish with additional cigarettes. While we do not find robust evidence that smoking intensity affects cognition, the insignificant effects may be due to the small size of the samples used in this analysis. Next, survivorship bias is assessed by estimating separate

models for persons aged 50–65 years and those aged 66–85 years (Table 8). In the case of the full sample (never, former and current smokers), there is some evidence of a downward bias for the 66–85 age group relative to the 50–65 age group. For the older age group, smoking has a larger negative effect on self memory, serial 7 and mental status and a smaller positive effect on past memory and total recall. In the case of the ever smoker sample, smoking has a larger positive effect on self memory and total recall and a smaller effect on past memory and serial 7 for the older age group. Despite these findings, the estimated effects are not substantially different in magnitude across the two age groups. This suggests that although survivorship bias may be a concern it does not fully account for the estimated results. As mentioned above, another concern with our analysis may be that the exclusion of proxy respondents from our sample biases the

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

8

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx

Table 6 Assessing the role of time varying factors: FE models.

Current smoker Retired Log income SRH = very good SRH = good SRH = fair SRH = poor Count of ADLs Count of IADLs High BP Diabetes Heart problems Stroke Psych. problems CESD Age Age squared Married/partnered

(1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

0.0796*** (0.0169) 0.0121 (0.0103) 0.0096* (0.0057) 0.0950*** (0.0137) 0.1783*** (0.0154) 0.3043*** (0.0185) 0.3758*** (0.0252) 0.0154** (0.0073) 0.1075*** (0.0134) 0.0005 (0.0148) 0.0117 (0.0208) 0.0150 (0.0181) 0.1097*** (0.0292) 0.0613** (0.0253) 0.0227*** (0.0025) 0.0068 (0.0131) 0.0000 (0.0001) 0.0463*** (0.0177)

0.0284*** (0.0091) 0.0170*** (0.0059) 0.0049 (0.0032) 0.0169** (0.0071) 0.0356*** (0.0080) 0.0867*** (0.0097) 0.1403*** (0.0135) 0.0248*** (0.0041) 0.0302*** (0.0081) 0.0007 (0.0085) 0.0094 (0.0122) 0.0069 (0.0109) 0.0318* (0.0177) 0.0309* (0.0164) 0.0210*** (0.0016) 0.0032 (0.0075) 0.0000 (0.0000) 0.0070 (0.0101)

2.7349*** (0.6852) 55,656 12,334

2.3258*** (0.3912) 55,656 12,334

0.0010 (0.0255) 0.0177 (0.0163) 0.0124 (0.0091) 0.0200 (0.0204) 0.0161 (0.0232) 0.0031 (0.0270) 0.0159 (0.0358) 0.0381*** (0.0112) 0.1464*** (0.0207) 0.0029 (0.0242) 0.0240 (0.0319) 0.0184 (0.0285) 0.1869*** (0.0442) 0.0106 (0.0398) 0.0135*** (0.0040) 0.1169*** (0.0204) 0.0010*** (0.0001) 0.0298 (0.0273) 0.0494* (0.0272) 0.0908 (1.0856) 55,656 12,334

0.1664*** (0.0576) 0.1388*** (0.0363) 0.0441** (0.0190) 0.0812 (0.0499) 0.1738*** (0.0561) 0.1980*** (0.0640) 0.3414*** (0.0811) 0.1024*** (0.0229) 0.3300*** (0.0430) 0.0419 (0.0548) 0.0284 (0.0726) 0.1131* (0.0647) 0.5324*** (0.1008) 0.1085 (0.0902) 0.0236*** (0.0088) 0.4483*** (0.0465) 0.0033*** (0.0002) 0.0569 (0.0603) 0.1219* (0.0644) 4.5848* (2.4745) 55,656 12,334

0.0266 (0.0590) 0.0405 (0.0323) 0.0138 (0.0218) 0.0578 (0.0400) 0.0560 (0.0446) 0.0708 (0.0528) 0.0165 (0.0696) 0.1117*** (0.0222) 0.6308*** (0.0487) 0.0073 (0.0488) 0.0036 (0.0618) 0.0235 (0.0531) 0.4212*** (0.0876) 0.1943** (0.0916) 0.0027 (0.0081) 0.6616*** (0.0560) 0.0044*** (0.0003) 0.0472 (0.0526) 0.1685*** (0.0371) 11.1086*** (2.7238) 34,802 11,024

0.1510 (0.1207) 0.1980*** (0.0656) 0.1013** (0.0426) 0.0798 (0.0860) 0.1931** (0.0954) 0.1940* (0.1080) 0.4409*** (0.1365) 0.2406*** (0.0386) 0.9198*** (0.0809) 0.1212 (0.1014) 0.2094 (0.1292) 0.1633 (0.1081) 0.9393*** (0.1754) 0.5523*** (0.1796) 0.0074 (0.0155) 1.0813*** (0.1097) 0.0073*** (0.0006) 0.0625 (0.1021) 0.1657** (0.0733) 16.5162*** (5.4406) 34,802 11,024

Prior exposure Constant N Persons

Regressions also include year and census region of residence dummies. Robust standard errors presented in parentheses are clustered at the individual level. * p < 0.10. ** p < 0.05. *** p < 0.01.

Table 7 Impact of smoking intensity among current smokers on cognition: FE models.

Panel A: Basic covariates # Of cigarettes # Of cigarettes squared N Persons Panel B: Additional covariates # Of cigarettes # Of cigarettes squared N Persons

(1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

0.0014 (0.0019) 0.0000 (0.0000) 13,905 4,168

0.0006 (0.0010) 0.0000 (0.0000) 13,905 4,168

0.0010 (0.0026) 0.0000 (0.0000) 13,905 4,168

0.0109** (0.0055) 0.0002** (0.0001) 13,905 4,168

0.0071 (0.0064) 0.0001 (0.0001) 6,675 3,139

0.0108 (0.0123) 0.0002 (0.0002) 6,675 3,139

0.0009 (0.0018) 0.0000 (0.0000) 13,683 4,149

0.0010 (0.0010) 0.0000 (0.0000) 13,683 4,149

0.0011 (0.0027) 0.0000 (0.0000) 13,683 4,149

0.0105* (0.0057) 0.0002** (0.0001) 13,683 4,149

0.0073 (0.0063) 0.0001 (0.0001) 6,594 3,101

0.0063 (0.0120) 0.0001 (0.0002) 6,594 3,101

Robust standard errors presented in parentheses are clustered at the individual level. Regressions also include age squared, prior exposure, and dummies for married or partnered status, year and census region of residence. * p < 0.10, ** p < 0.05.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

9

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx Table 8 Addressing survivorship bias: FE models. (1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

Sample: All persons aged 50 to 65 (basic covariates) Current Smoker 0.2766 (0.4278) Former Smoker 0.3607 (0.4282) N 44,149 Persons 13,070

0.1013 (0.0976) 0.0454 (0.0981) 44,149 13,070

0.1244 (0.2404) 0.1433 (0.2413) 44,149 13,070

1.1030 (1.0254) 0.8920 (1.0282) 44,149 13,070

1.9173 (1.6939) 1.9555 (1.7074) 8,473 8,178

1.7849 (3.4061) 2.4764 (3.4362) 8,473 8,178

Sample: All persons aged 66 to 85 (basic covariates) Current Smoker 1.1101*** (0.0158) Former Smoker 1.2152*** (0.0235) N 48,333 Persons 13,906

0.0564*** (0.0086) 0.0350*** (0.0125) 48,333 13,906

0.4024*** (0.0247) 0.4168*** (0.0361) 48,333 13,906

0.9722*** (0.0574) 0.7507*** (0.0850) 48,333 13,906

0.9911*** (0.0370) 1.0041*** (0.0522) 48,286 13,903

0.0159 (0.0745) 0.2538** (0.1062) 48,286 13,903

Sample: Ever smokers aged 50 to 65 (basic covariates) 0.0535*** Current smoker 0.0794*** (0.0231) (0.0126) N 26,150 26,150 Persons 7,767 7,767

0.0115 (0.0345) 26,150 7,767

0.2015** (0.0788) 26,150 7,767

0.0000 (0.3339) 4,986 4,806

0.4129 (0.6998) 4,986 4,806

Sample: Ever smokers aged 66 to 85 (basic covariates) Current Smoker 0.0995*** 0.0198 (0.0296) (0.0153) N 27,975 27,975 Persons 8,185 8,185

0.0049 (0.0443) 27,975 8,185

0.2101** (0.1031) 27,975 8,185

0.0017 (0.0646) 27,952 8,182

0.2147* (0.1300) 27,952 8,182

Sample: Ever smokers aged 50 to 65 (additional covariates) Current Smoker 0.0692*** 0.0394*** (0.0232) (0.0127) N 25,320 25,320 Persons 7,697 7,697

0.0152 (0.0355) 25,320 7,697

0.1735** (0.0808) 25,320 7,697

0.4386 (0.3598) 4,826 4,657

0.9721 (0.8532) 4,826 4,657

Sample: Ever smokers aged 66 to 85 (additional covariates) Current Smoker 0.0884*** 0.0121 (0.0295) (0.0153) N 27,534 27,534 Persons 8,119 8,119

0.0038 (0.0448) 27,534 8,119

0.1961* (0.1050) 27,534 8,119

0.0182 (0.0640) 27,512 8,116

0.1807 (0.1307) 27,512 8,116

Robust standard errors presented in parentheses are clustered at the individual level. * p < 0.10, ** p < 0.05, *** p < 0.01.

Table 9 Smoking rates by survey year and proxy status. Selfrespondents (%) Panel A: Ever smoker 1998 59.04 2000 58.69 2002 58.16 2004 57.58 2006 57.08 2008 56.74

Proxy respondents (%)

Difference (%)

p-Value from a test of differences

58.88 57.99 57.67 56.08 55.15 55.82

0.16 0.70 0.49 1.50 1.93 0.92

0.8947 0.5525 0.6827 0.2320 0.1920 0.5515

Panel B: Current smoker, conditional on being an ever smoker 1998 27.85 26.39 1.46 0.2945 2000 25.56 23.45 2.11 0.1192 2002 23.58 23.08 0.50 0.7091 2004 25.50 21.48 4.02 0.0061 2006 23.75 16.42 7.33 <0.001 2008 22.97 15.31 7.66 <0.001

estimates. To assess the extent of such bias, Table 9 presents the smoking rates of self-respondents and of persons represented by a proxy for each year in our sample. Panel A presents the percentage of individuals who report being an ever smoker. Panel B presents the percentage of individuals who report being a current smoker, conditional on being an ever smoker. The last column of

Table 9 presents the p-value from a test of significant differences in smoking rates by proxy status. We do not find any significant differences in the percentage of persons who report being an ever smoker, by proxy status. However, we find a different pattern when we examine the percentage of current smokers conditional on being an ever smoker. In the 1998 through 2002 waves, there are no significant differences by proxy status. However, in the 2004 through 2008 waves, individuals represented by a proxy are significantly less likely to report being a current smoker compared to self-respondents. If smoking reduces cognition then this would suggest that our estimates are biased downwards due to the exclusion of proxy respondents, and could be viewed as a lower bound of the true effect. Thus, while we cannot conclusively rule out bias due to exclusion of proxy respondents for the regressions using the ever smoker sample, the direction of this bias should result in more conservative estimates. In Table 10, we evaluate the ‘‘sick quitter’’ phenomenon, which is the finding that individuals who have quit smoking recently often report worse health and incur higher medical expenditures than smokers, likely because quitting may be in response to a health shock. To assess the extent to which our results may be driven by ‘‘sick quitters’’ we identify persons who quit smoking within the past 3 years based on a HRS question that asks former smokers to report how long ago they stopped smoking. We exclude these recent quitters from our sample and re-estimate the fixed

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

10

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx

Table 10 Evaluating the sick quitter phenomenon: FE models. (1) Self memory

(2) Past memory

(3) Serial 7

(4) Total recall

(5) Mental status

(6) Total cognition

Panel A: Full sample (basic covariates) Current smoker 0.3183 (0.3313) Former smoker 0.4240 (0.3316) N 96,711 Persons 21,226

0.1435* (0.0783) 0.1010 (0.0787) 96,711 21,226

0.0657 (0.2813) 0.0419 (0.2814) 96,711 21,226

0.8962 (0.7755) 0.6913 (0.7783) 96,711 21,226

2.5101*** (0.8669) 2.5024*** (0.8698) 61,049 18,923

2.0676 (1.6084) 2.2911 (1.6131) 61,049 18,923

Panel B: Full sample (additional covariates) Current smoker 0.3688 (0.3241) Former smoker 0.4575 (0.3244) N 95,650 Persons 21,179

0.1263* (0.0669) 0.0975 (0.0673) 95,650 21,179

0.1746 (0.3134) 0.1635 (0.3134) 95,650 21,179

0.8847 (0.7570) 0.7252 (0.7597) 95,650 21,179

2.2508*** (0.8359) 2.1985*** (0.8388) 60,517 18,797

1.6023 (1.7047) 1.7317 (1.7089) 60,517 18,797

Panel C: Ever smoker sample (basic covariates) Current Smoker 0.1013*** (0.0176) N 55,698 Persons 12,360

0.0402*** (0.0095) 55,698 12,360

0.0159 (0.0259) 55,698 12,360

0.1944*** (0.0601) 55,698 12,360

0.0173 (0.0604) 34,574 10,908

0.1930 (0.1256) 34,574 10,908

Panel D: Ever smoker sample (additional covariates) Current Smoker 0.0864*** (0.0174) N 55,070 Persons 12,332

0.0274*** (0.0095) 55,070 12,332

0.0052 (0.0261) 55,070 12,332

0.1508** (0.0603) 55,070 12,332

0.0555 (0.0593) 34,279 10,827

0.1104 (0.1249) 34,279 10,827

Robust standard errors presented in parentheses are clustered at the individual level. * p < 0.10, ** p < 0.05, *** p < 0.01.

effects models. A comparison of the estimates in Table 10 with the estimates in Tables 3, 5 and 6 shows that our results are robust to excluding recent quitters. Discussion Using data from the HRS and employing fixed effects models, we find that the negative association between smoking and cognitive function identified by prior literature can be partly attributable to unobserved differences between smokers and non-smokers. While current and former smokers exhibit worse cognition than never smokers in cross-sectional models, this negative relationship does not hold for most measures of cognition when we include individual fixed effects. Moreover, we find no evidence that smoking cessation improves cognition once unobserved person specific factors are accounted for. Using the ever smoker sample, we find that current smokers perform better than former smokers on some measures of cognition such as word recall. While this is consistent with the hypothesis that nicotine improves attention and inhibits the formation of senile plaques in the brain, our estimates do not fully account for biases due to the exclusion of proxy respondents, unobserved time-varying factors or differential mortality. Thus, further research is required to understand whether nicotine may indeed have protective effects. Our study is the first to use a nationally representative sample to identify the impact of smoking on cognition. Given the aging population in the U.S. and significant health care costs associated with Alzheimer’s and other forms of dementia, there is considerable interest in identifying factors that might slow cognitive impairment among older adults. Our results suggest that individual health behaviors could be successful in this; however, further research is required to identify healthy activities that may aid brain health without the harmful side effects of cigarette use. Given the well-established and widely documented harmful health effects of tobacco use, one must be cautious in evaluating the

implications of our findings. The main contribution of our study is to highlight the role of confounding factors in explaining the conflicting results found in existing literature. Our analysis also highlights the need for further medical research into the potential role of nicotine in developing new and safe drug therapy for preventing Alzheimer’s disease and other dementias. An important limitation of our econometric analysis is that fixed effects models do not account for unobserved time varying factors that might bias the estimated results. The information available in the HRS allows us to adjust for an important but still limited set of time varying factors. Future work should explore other data sets or estimation strategies that could address this limitation.

Acknowledgment The authors are grateful to participants at the 2012 Southern Economic Association Conference for helpful comments and suggestions.

References Almeida, O.P., Hulse, G.K., et al., 2002. Smoking as a risk factor for Alzheimer’s disease: contrasting evidence from a systematic review of case–control and cohort studies. Addiction 97 (1), 15–28. Almeida, O.P., Garrido, G.J., et al., 2011. 24-Month effect of smoking cessation on cognitive function and brain structure in later life. NeuroImage 55 (4), 1480– 1489. Alzheimer’s Association Report, 2012. 2012 Alzheimer’s disease facts and figures. Alzheimer’s Dementia 8 (2), 131–168. Alzheimer’s Disease Education & Referral (ADEAR) Center. (2011). ‘‘Alzheimer’s Disease Fact Sheet’’. Retrieved July 25, 2013. Anstey, K.J., von Sanden, C., et al., 2007. Smoking as a risk factor for dementia and cognitive decline: a meta-analysis of prospective studies. Am. J. Epidemiol. 166 (4), 367–378. Banks, J., Mazzonna, F., 2012. The effect of education on old age cognitive abilities: evidence from a regression discontinuity design. Econ. J. 122 (560), 418–448. Bonsang, E., Adam, S., et al., 2012. Does retirement affect cognitive functioning? J. Health Econ. 31 (3), 490–501.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001

P. Ayyagari, A. Sikora Kessler / The Journal of the Economics of Ageing xxx (2015) xxx–xxx Case, A., Paxson, C., 2009. Early life health and cognitive function in old age. Am. Econ. Rev. 99 (2), 104–109. de Walque, D., 2007. Does education affect smoking behaviors?: evidence using the Vietnam draft as an instrument for college education. J. Health Econ. 26 (5), 877–895. de Walque, D., 2010. Education, information, and smoking decisions. J. Human Resour. 45 (3), 682–717. Falba, T., 2005. Health events and the smoking cessation of middle aged Americans. J. Behav. Med. 28 (1), 21–33. Farrer, L.A., Cupples, L.A., et al., 1997. Effects of age, sex, and ethnicity on the association between apolipoprotein E genotype and Alzheimer disease. J. Am. Med. Assoc. 278 (16), 1349–1356. Fisher, G., Hassan, H., et al. (2012). Health and Retirement Study—Imputation of cognitive functioning measures: 1992–2010 early release. Ann Arbor, MI, Survey Research Center. Retrieved March 5: 2012. Fratiglioni, L., Wang, H.-X., 2000. Smoking and Parkinson’s and Alzheimer’s disease: review of the epidemiological studies. Behav. Brain Res. 113 (1), 117–120. Guven, C., Lee, W.S., 2013. Height and cognitive function at older ages: is height a useful summary of early childhood. Health Econ. 22 (2), 224–233. Henkens, K., van Solinge, H., et al., 2008. Effects of retirement voluntariness on changes in smoking, drinking and physical activity among Dutch older workers. Eur. J. Public Health 18 (6), 644–649. Herzog, A. Wallace, R. B. (1997). Measures of cognitive functioning in the AHEAD Study. J. Gerontol. Series B: Psychol. Sci. Soc. Sci. 52(Special Issue): 37. Hjelm, C., Dahl, A., et al., 2012. The influence of heart failure on longitudinal changes in cognition among individuals 80 years of age and older. J. Clin. Nurs. 21 (7–8), 994–1003. Jha, P., Ramasundarahettige, C., et al., 2013. 21st-Century hazards of smoking and benefits of cessation in the United States. N. Engl. J. Med. 368 (4), 341–350. Jorm, A., Jacomb, P., 1989. The informant questionnaire on cognitive decline in the elderly (IQCODE): socio-demographic correlates, reliability, validity and some norms. Psychol. Med. 19 (4), 1015–1022. Keenan, P.S., 2009. Smoking and weight change after new health diagnoses in older adults. Arch. Intern. Med. 169 (3), 237. Kenkel, D.S., 1991. Health behavior, health knowledge, and schooling. J. Polit. Econ., 287–305 Khwaja, A., Sloan, F., et al., 2006. Learning about individual risk and the decision to smoke. Int. J. Ind. Organ. 24 (4), 683–699. Lacey, R.E., Cable, N., et al., 2011. Childhood socio-economic position and adult smoking: are childhood psychosocial factors important? Evidence from a British birth cohort. Eur. J. Public Health 21 (6), 725–731. Maurer, J., 2010. Height, education and later-life cognition in Latin America and the Caribbean. Econ. Human Biol. 8 (2), 168–176. Mazzonna, F., Peracchi, F., 2012. Ageing, cognitive abilities and retirement. Eur. Econ. Rev. 56 (4), 691–710. Murphy, S., Xu, J., et al. (2013). Final Data for 2010. National Vital Statistics Reports. Hyattsville, MD, National Center for Health Statistics. 61(4).

11

Newhouse, P., Kellar, K., et al., 2012. Nicotine treatment of mild cognitive impairment A 6-month double-blind pilot clinical trial. Neurology 78 (2), 91– 101. Ofstedal, M.B., Fisher, G.G., et al., 2005. Documentation of cognitive functioning measures in the Health and Retirement Study. University of Michigan, Ann Arbor, MI. Peters, R., Poulter, R., et al., 2008. Smoking, dementia and cognitive decline in the elderly, a systematic review. BMC Geriatr. 8 (1), 36. Qiu, C., Kivipelto, M., et al., 2009. Epidemiology of Alzheimer’s disease: occurrence, determinants, and strategies toward intervention. Dial. Clin. Neurosci. 11 (2), 111. Rezvani, A.H., Levin, E.D., 2001. Cognitive effects of nicotine. Biol. Psychiatry 49 (3), 258–267. Rodgers, W.L., Ofstedal, M.B., et al., 2003. Trends in scores on tests of cognitive ability in the elderly US population, 1993–2000. J. Gerontol. Series B: Psychol. Sci. Soc. Sci. 58 (6), S338–S346. Roe, C.M., Xiong, C., et al., 2007. Education and Alzheimer disease without dementia support for the cognitive reserve hypothesis. Neurology 68 (3), 223–228. Rohwedder, S., Willis, R.J., 2010. Mental retirement. J. Econ. Perspect. 24 (1), 119– 138. Salomon, A.R., Marcinowski, K.J., et al., 1996. Nicotine inhibits amyloid formation by the b-peptide . Biochemistry 35 (42), 13568–13578. Savva, G.M., Stephan, B.C., 2010. Epidemiological studies of the effect of stroke on incident dementia a systematic review. Stroke 41 (1), e41–e46. Srivareerat, M., Tran, T.T., et al., 2011. Chronic nicotine restores normal Ab levels and prevents short-term memory and E-LTP impairment in Ab rat model of Alzheimer’s disease. Neurobiol. Aging 32 (5), 834–844. Stern, Y., Gurland, B., et al., 1994. Influence of education and occupation on the incidence of Alzheimer’s disease. J. Am. Med. Assoc. 271 (13), 1004–1010. Taylor, D.H., Hasselblad, V., et al., 2002. Benefits of smoking cessation for longevity. Am. J. Public Health 92 (6), 990–996. Tyas, S.L., White, L.R., et al., 2003. Mid-life smoking and late-life dementia: the Honolulu-Asia Aging Study. Neurobiol. Aging 24 (4), 589–596. Ulrich, J., Johannson-Locher, G., et al., 1997. Does smoking protect from Alzheimer’s disease? Alzheimer-type changes in 301 unselected brains from patients with known smoking history. Acta Neuropathol. 94 (5), 450–454. Whitmer, R., Sidney, S., et al., 2005. Midlife cardiovascular risk factors and risk of dementia in late life. Neurology 64 (2), 277–281. Wilson, A.L., Langley, L.K., et al., 1995. Nicotine patches in Alzheimer’s disease: pilot study on learning, memory, and safety. Pharmacol. Biochem. Behav. 51 (2–3), 509–514. Yamada, M., Kasagi, F., et al., 2003. Association between dementia and midlife risk factors: the radiation effects research foundation adult health study. J. Am. Geriatr. Soc. 51 (3), 410–414. Zanardi, A., Leo, G., et al., 2002. Nicotine and neurodegeneration in ageing. Toxicol. Lett. 127 (1–3), 207–215.

Please cite this article in press as: Ayyagari, P., Sikora Kessler, A.. The Journal of the Economics of Ageing (2015), http://dx.doi.org/10.1016/ j.jeoa.2015.06.001