Ch 4 Understanding Change in Aging Research

parametric

- It has to do with the underlying distribution of the variable not matter what the numbers look like. - Parametric statistics are used when the underlying distribution of the variable is continuous. - Nonparametrics are used when the underlying distributions of variables are categorical.

Mean-Level Change: Summary of Common Longitudinal Statistics

- paired t-test: examines change from Time 1 to Time 2 in two groups - Repeated-measures MANOVA: examines change across multiple time periods in multiple groups

Dose-response curve: Experimental Designs

- plots the symptoms against amount of medication, or it can also refer to the relationship between toxic substances and symptoms. - Think Flint and lead or opiod overdoses. How does the dose-response idea apply to those situations. - Classic finding is called a dose-response effect, and refers to a linear relationship

Principal axes

- seek to identify different dimensions within a larger construct. - This is the technique used by personality researchers to identify the number of dimensions are needed to describe personality.

Principal components

- seek to identify patterns of similarities in the items that most closely track a construct. > For example, I recent collected about 20 variables from US Census records that characterize census blocks or entire counties. - Principal components boilded it down to two variables, That's a lot easier to work with.

Parametric statistics

- such as t-tests and analyses of variances (ANOVAs) are used when the dependent variables are continuous. - Gender is not continuous, so it really makes no sense to report the mean for gender.

moderating effect

- the effect of one variable on another one may be moderated: that is, it may be conditional on some other variable such as context, age, or gender. - For example, individuals high in hostility may have elevated blood pressure levels primarily in situations in which there is interpersonal conflict. > Under normal conditions, their blood pressure may be similar to individuals who are low in hostility.

Two-Point Designs: Statistics That Evaluate Change

- uses baseline data to predict some outcome - Simplest way of doing this is to use either multiple regression or logistic regression - Residualized regressions are hierarchal regressions in which the Time-1 dependent variable is entered first and thus covaried out, and the other variables are used to predict the adjusted Time-2 scores

ppt

(textbook flashcard 36)

A good example of the importance of random assignment: Experimental Designs

- A good example of the importance of random assignment is the effect of combination estrogen/progesterone hormone replacement therapy (HRT) on health. - Early studies showed that women who used HRT tended to be healthier than those who did not. > However, women on HRT also tend to be better educated and more health conscious. > Only when they were randomly assigned to the group receiving HRT or the control group could researchers establish that HRT actually resulted in slightly increased risks of heart disease and breast cancer (Rossouw et al., 2002): at least among older women

partial correlation

- A partial correlation controls for one variable a when correlating two other variables (b with c). - For example, education, a, and income, b, are highly intercorrelated, and both are correlated with health, c. - To determine whether income has an effect independent of education, one would partial out education. - Canonical correlations simply examine whether two sets of variables are correlated.

regression equation

- A regression equation is basically a series of partial correlations. - It does much the same thing but with several variables so that one can examine the contribution of each variable independent of the others in the equation. - Thus, one could construct an equation that examined the independent contributions of occupation, education, income, and job satisfaction to health outcomes. - Regression can be used to decide which variables best predict a given outcome (for a very comprehensive online tutorial on regression, see Econometrics Laboratory, University of California at Berkeley, 1999). - Structural equation modeling (SEM) simultaneously estimates the relation: ships among several variables and will be discussed later in this chapter.

A second way to measure change at the individual level: ipstative change

- A second way of examining change at the individual level is to calculate change or difference scores (e.g., T2 - T1). - Several decades ago, Cronbach and Furby (1970) argued strenuously against the use of change scores because they did not take into account differences in where individuals started. > For example, a two-point rise in body temperature from 98F to 100F does not signify very much, but a change from 104F to 106F might well prove fatal. > Cronbach and Furby argued that researchers instead should use residualized regressions to assess change (see below). - However, Rogosa (1988) argued for the utility of change scores, and sometimes simple change scores can prove very informative. > For example, Spiro et al. (2000) showed how different methods of assessing change and stability in the same sample can yield dramatically different results. > Over a 5-year period, Minnesota Multiphasic Personality Inventory 2 (MMPI-2) subscale scores show great correlational stability, usually about .8. > The standard interpretation would be that personality traits are very stable; however, change scores showed that 15% to 20% of the sample changed one standard deviation or more on each scale, which was clinically significant. > Examining change over the various subscales showed that 70% of the sample had a clinically significant change on at least one of the subscales.

case control: Quasi-Experimental Designs

- A special type of quasi-experimental design sometimes used in health research is called case control (Norman & Streiner, 2000). - A quasi-experimental design may not be practical when studying a relatively rare phenomenon; huge samples would be needed to identify enough cases to analyze. > In this instance, case control designs can be useful. > For example, if there is a cluster of a few deaths in a particular area, one can identify who died (cases), then identify a group of living people similar to those who died (control), and then compare the two groups to see what difference between the two might account for the deaths.

Cross-Sectional Versus Longitudinal Designs

- Age differences are studied using cross-sectional designs, which compare different age groups at one point in time - Age-related changes are examined using longitudinal designs, which follow people over time - Botwinick (1977) in a classic study using the Duke Longitudinal Study of Aging (DLSA) sample, confirmed that cross-sectional differences in IQ showed an apparent decline with age, but a longitudinal study on the same sample show an increase in IQ with age. - As a side note, most resarchers actually avoid IQ because IQ is age normed. > your score depends upon how your score compares to those of other people your age. > an IQ means of 100 means you scored at exactly the mean. - most of the time researchers use the raw score (the number of items correct) when doing cross-sectional or sequential studies. - For longitudinal studies, there are generally fewer respondents at each assessment due to deaths, dropouts, and individuals who have move and been lost to follow-up > Excluding individuals with even one missing data point often results in losing substantial proportions of the sample, and researchers have developed innovative ways to estimate missing data > there is also an issue called terminal drop - Longitudinal designs confound aging effects with cohort effects, as one cohort may change in a way that other cohorts do not - Longitudinal studies also confound period (historical or time of measurement) effects. - Because longitudinal work is so difficult to conduct, researchers starting with K. W. Schaie in 1965 developed sequential designs to allow us to examine our data for age, cohort, and period. Then there's test-retest effects.

Relative Change: statistics for assessing change

- Another way of assessing change and stability is to simply correlate two assessments across time. > This indicates the extent to which individuals in the sample maintain their relative rank order. - Figure 4.2A illustrates one pattern of correlational stability. > Let us assume that this figure illustrates change in introversion with age. > At Time 1, the scores on introversion are normally distributed, and Jill has an introversion score of 7, which is higher than Jack's, who scores at the mean (5). > In turn, Jack has a higher score than Rachel (3). > By Time 2, the population as a whole has increased in introversion, but there is relative stability. > Everyone has increased by two points. Jill's extraversion score (9) still has a higher level than Jack's (7), who still has a higher level than Rachel (5). > Thus, correlations may indicate stability of individual differences even when there actually is a mean-level change occurring. - It is also possible to have no significant mean-level change across time while there is a change in rank order. - As Figure 4.2B indicates, the population mean can remain the same, but the individuals may change. > In this instance, Jill has become withdrawn following a series of negative life events and has become much more introverted than the other two. > Rachel is now working as a real estate agent and must become more extraverted so has raised her score from 3 at Time 1 to 7 at Time 2. > Jack has not changed; he either is in a rut or is simply happy the way he is. > The net mean change, however, is zero, so the population mean is stable, but there is correlational instability: there may be a negative correlation over time, as high introversion scores at Time 1 are associated with low ones at Time 2, and vice versa. - Many researchers assume that a significant correlation indicates stability, when, in truth, stability and change are always relative, and the magnitude of the correlation should be examined to assess the degree of stability. - In general, correlations over .8 indicate a high degree of stability, correlations of .4 to .6 indicate moderate stability, and .2 to .4 show weak stability (Cohen & Cohen, 1983). - One factor that affects the magnitude of the correlation is the internal reliability of the scale, which is a measure of how well the items in the scale intercorrelate. > If the items do not ''hang together'' very well, reliability may be low: meaning that it does not clearly assess the construct. - If the construct is not assessed very well, then the cross-time correlations may be lower than they should be. - Nunnally (1978) recommends calculations that correct cross-time correlations for the internal reliability of the scales. - Note that if scales are highly unreliable, these corrections may result in cross-time correlations that are greater than 1.0 (see Spiro, Butcher, Aldwin, & Levenson, 2000). - fig p 71

Cross Sectional vs. Longitudinal Designs: Comparison of the Different Designs

- Another way of categorizing research designs is whether or not they are cross- sectional (e.g., collect data at one point in time) or are longitudinal (e.g., collect data across at least two time points). - Longitudinal designs are particularly important for aging research

Structural Change: statistics for assessing change

- As mentioned earlier, factor analysis examines the structure or patterns in a series of correlations. > For example, scales assessing depression typically include items that assess negative affect (feeling blue), negative cognition (thinking that everything is hopeless), as well as physical symptoms (being exhausted). > Thus, factor analyzing such scales should yield three factors: affect, cognition, and symptoms. > In other words, the affect items should intercorrelate fairly strongly, as should the cognitive ones and physical ones. > A factor loading is the strength of the association between an item and a factor. > Thus, the items assessing affect should load more heavily on the affective factor, and so on. - Factor analysis can be used to identify the best items to assess the construct (principal components), or it can be used to determine if there are different dimensions or facets of a construct (principal axis). - Factor structures can change over time or vary in different samples or age groups if the meaning of the items changes. > For example, in a young person, the physical symptoms may be indicative of depression. In an older person, they may indicate depression, or they may also reflect either normal age-related changes in sleep patterns, physical illness, or side effects of medication. > If this is true, one would expect that the factor structure of depression measures in younger individuals might be different from those used with older individuals; for example, in a younger per- son, the physical symptoms may load on the affective factor. - It may be difficult to tell whether a change in the factor loadings is really significant or just varies by chance. - Confirmatory factor analysis (CFA) is used to assess structural change and stability and is a specific type of SEM (Bentler, 1998; Long, 1983; MacCallum & Austin, 2000). > SEM basically uses a type of factor analysis to ''purify'' the variance in a scale (i.e., make it more reliable) and then uses regression analysis to examine the relationship between these ''purified'' constructs. - CFA allows a researcher to determine whether the factor loadings for a construct at Time 1 (or in one age group) are significantly different from the factor loadings for a construct at Time 2 (or in a second age group). > Figure 4.3 illustrates this: Let's assume we have a depression measure consisting of six items: two affective, two cognitive, and two physical. > The indicator variables (or scale items) are represented by squares, and the latent construct (purified measure) of depression is represented by an oval. > The factor loadings between the indicator variables and the latent construct are symbolized by the Greek letter lambda (). > So 1 represents the strength of the association between affect 1 and depression, 2 between affect 2 and depression, and so on. > In Figure 4.3, the first six lambdas (1 through 6) represent the relationships among the items and the overall construct of depression in the younger group, and the second set of lambdas (7 through 12) indicates the same relationships in the older group. > If there is structural stability, the s should be comparable across the two models; that is, 1 should be similar to 6, 2 to 7, and so on. > To determine this, CFA contrasts two models (Byrne, Shavelson, & Muthen, 1989). > In Model 1, the Time 1 and Time 2 s are assumed not to be equal but are allowed to vary freely, and in Model 2, the s are assumed that they are equal and are fixed. > For each model, the degree to which the hypothesized model differs from the actual data is indicated by a 2: The smaller the 2, the better the model fits the actual data. > A very good model is one in which the 2 is not significant—that is, the observed model is not different from the predicted model. - CFA calculates the difference between the two 2s to determine whether the models are significantly different from each other. - If the factor structure is not stable, we assume that the 2 for Model 2 would be bigger than Model 1 because it is not a good fit to the data. If the 2 for the two models is not significantly different, then the freely estimated model is similar to (or replicates) the constrained model, so we would conclude that the factor structure is relatively stable across age groups. - CFA is a very complex technique, but its use is increasing in frequency. - CFA is very important because it tests a critical assumption in developmental research. - If the factor structure of a particular measure is not relatively stable across age groups or across time, you cannot use that measure to examine cross-sectional differences or change because the meaning of the measure varies by age. p 73

Cross-Lagged Panel Designs and Analyses: Statistics That Evaluate Change

- Can be used to determine the relative strength of each pathway - Can be done by computing two residualized regressions - Can be done by using SEM, which simultaneously estimates all pathways

STATISTICS FOR ASSESSING CHANGE

- Caspi and Bem (1990) categorized change and stability into four different types: 1) absolute or mean level 2) relative 3) structural 4) idiothetic or ipsative. - Absolute change refers to changes in mean level over time, whereas relative change refers to shifts in rank order. > Note that these two types of change are independent of each other. > It is possible to have stable means across time but changes in rank order and vice versa. - Structural change refers to whether or not the factor structure (pattern of relationships) is similar across age or time. > Note also that all of these types of change refer to changes in populations. - In contrast, idiothetic or ipsative change refers to changes in individuals, and different types of statistics are required for this type of assessment. - These types of longitudinal statistics are summarized in Table 4.3. p 69

Comparison of the Different Designs: pros and cons

- Clinical studies, especially those using case control designs, may overestimate the effect size - Large surveys using quasi-experimental designs often underestimate the effect size - Experimental designs may not find an effect because the sample is too small or the treatment was not well designed. - Correlational designs often use convenience samples whose generalizability may be limited and certainly cannot address causality. - Psychologists are not always good at demography or don't have the resources to recruit a truly random sample. > Not even the CDC or the census bureau are perfect. - Longitudinal designs are particularly important aging research and particularly difficult to pull off because people don't always return for follow up sessions.

Cohort-Sequential Designs: age-related designs

- Cohort-sequential designs follow two or more cohorts over the same ages but confound time (see panel A, Table 4.2). - Members of the first cohort illustrated in Table 4.2, born in 1947, were 22 years old at the first time of measurement in 1969 and 32 years old in 1979. > Members of the second cohort, born 10 years later in 1957, were also assessed at ages 22 and 32, but their assessments occurred in 1979 and 1989. > Thus, we have two cohorts (1947 and 1957), two ages (22 and 32), but three time periods (1969, 1979, and 1989). > Therefore, period is confounded because the cells are unbalanced. > The 1957 cohort had no measurements in 1969, and the 1989 assessment of the 1947 cohort was not used because the age was out of range (age 42); therefore, cohort-sequential designs can only test age and cohort effects, not period effects. - For a phenomenon to be an age effect, both cohorts must show similar (preferably identical) changes. - Parker and Aldwin (1997) found that the 1947 cohort, who graduated in 1969, increased in mastery from ages 22 to 32. > To determine if this was an age or cohort effect, they also studied the 1957 cohort, who graduated in 1979, and also found that this second cohort increased in mastery nearly the identical amount from ages 22 to 32; however, it was entirely possible that this was a period effect: that there was a general increase in mastery over this time period. > Therefore, Parker and Aldwin also conducted a cross-sequential study.

Structural Change: Summary of Common Longitudinal Statistics

- Confirmatory factor analysis: examines whether the pattern of relationships varies across time or groups

Experimental Designs

- Experiment is defined as "a test under controlled conditions that is made to demonstrate a known truth, examine the validity of a hypothesis, or determine the efficacy of something previously untried" - In an experimental design, random assignment to groups is absolutely critical to determine whether or not a given manipulation or treatment works - The text gives the example is the effect of a treatment that consistent of giving women a combination of estrogen/progesterone hormone replacement therapy (HRT) on health. - In a treatment study, the researcher must decide on the appropriate control group. > This is not always obvious but in medical settings the decision is often made that the control group must at least think that they are receiving a treatment. Why?

Observational Designs example

- For example, suppose we re interested in the public perception of driverless vehicles. - If I was assigned that task, I wouldn't know where to begin. - There are lots is issues but which ones are central to the question of will people accept them and actually purchase them. - As someone interested in aging I might want to include concerns that are of particular concern to people over 70 or 80. > Could these cars allow them to stay mobile for a loner period of time? > Would the technology frighten them. > How about parents of young children, would they feel safe? > Dozens of focus groups are being run by automakers around the world .

testing IQ (cross sectional vs. longitudinal designs): confounding age and cohort effects

- Cross-sectional and longitudinal designs can yield varying results. - Finding age differences in a phenomenon between different age groups does not mean that there are age-related changes. - Nearly every cross-sectional study finds that older people have lower IQ levels; however, we know that there is a cohort effect of education (older people tend to have less education than younger people), which can affect IQ levels. - In a classic study using the Duke Longitudinal Study of Aging (DLSA) sample, confirmed that cross-sectional differences in IQ showed an apparent decline with age, but he also conducted a longitudinal study on the same sample, which showed an increase in IQ with age. > One could conclude that the apparent age effect was actually a cohort effect, as cross-sectional studies confound age effects with those due to cohort. > Actually, neither of these findings was accurate. > The cross-sectional analysis did confound age with education, but the longitudinal analysis had a problem with survivor effects. - In longitudinal studies, there are generally fewer respondents at each assessment due to deaths, dropouts, and individuals who have moved and been lost to follow-up. - Typically, poorer functioning individuals tend to drop out of longitudinal studies, and the ones who remain are generally healthier, wealthier, and better educated. - Thus, in the longitudinal analysis of the DLSA sample, IQ appeared to increase with age but only because there was a more select sample at each time point. - When only data from those individuals who responded at all time points were included in the analysis, then IQ was much more stable, decreasing only after age 60 or so. - However, excluding individuals who have missing data still entails selection effects.

Relative Change: Summary of Common Longitudinal Statistics

- Cross-time correlation: examines association of two or more variables across time - Residualized regression: identifies predictors of change from Time 1 to Time 2 - Proportional-hazards mode: identifies predictors of the occurrence of an event over time

Nonparametric and parametric statistics

- Different statistics are used depending on whether or not the outcome variable is categorical or continuous. - Nonparametric statistics such as chi-square (2) analysis are used if both the independent (predictor) and dependent (outcome) variables are categorical. - Parametric statistics such as t-tests and analyses of variances (ANOVAs) are used when the dependent variables are continuous. > They compare the means between groups depending on whether there are two or more than two groups, respectively.

continuous vs. categorical variables

- Different statistics are used to answer these two questions depending on the type of data available (continuous or categorical) and, in part, which statistics are in vogue in different fields. - Continuous variables are those that form a series, such as age or number of symptoms, whereas categorical variables refer to groups that cannot be added together, such as where someone was born or type of illness. > In health- related research, when a study describes mean differences between the groups, it is assessing a continuous variable (e.g., number of symptoms) and describing average differences between the groups. - When a study discusses the probability of developing a health problem (odds or risk ratios), the dependent variable is categorical.

Direct effects

- Direct effects indicate that an independent variable, x, has a direct (or causal) impact on the dependent variable, y. > For example, hostility is associated with high blood pressure and it is possible that this is a direct effect

Direct, Mediating, and Moderating Effects: Special Design Issues in Health and Aging Research

- Direct effects indicate that and independent variable, X, has a direct (or causal) impact on the dependent variable, Y - Mediated effects occur when there is an association between two variables, but the effect is actually mediated by a third variable - Indirect effects occur when there is no direct relationship between a predictor and an outcome variable, but the predictor influences a third variable, which in turn influences the outcome - Moderated effects may be conditional on some other variable such as context, age, or gender

Correlational Designs

- Does not divide people in to groups but instead simply examines the association among groups of variables - Directionality: which variables influences the others is not something we can see with correlational designs. - Purpose is to determine the relationship between two variables, as well as how significant and strong the relationship is. - Positive relationship occurs when two variables increase or decrease together - Negative relationship occurs when one variable increases as the other decreases - Correlational studies can use a variety of different statistics depending on the nature of the variables and how they are measured. But the statistics do not change the underlying nature of the research.

Cross-Lagged Panel Designs and Analyses: STATISTICS THAT PREDICT CHANGE

- Even with longitudinal data, causal directionality can sometimes be questionable. > For example, we know that stressful life events can predict depressive symptoms, but people who are depressed may also be exposed to (or report) more life events. - A cross-lagged panel design can be used to determine the relative strength of each pathway. Figure 4.5 depicts this cross-lagged design (King, Taft, & King, 2006). - There are two ways of analyzing cross-lagged panel models: 1) to compute two residualized regressions. > The first predicts Time 2 stress, controlling for Time 1 stress (13) and examines the independent contribution of symptoms at Time 2 to stress at Time 2 (23). > The second equation predicts Time 2 depressive symptoms, controlling for Time 1 symptoms (24) and then entering Time 1 stress (14). > The relative sizes of the cross-lagged betas (24 vs. 14) reveal which path is stronger and therefore provide a clue as to causal directionality. 2) Another way of doing this analysis is to use SEM, which simultaneously estimates all pathways. > This can provide the cross-lagged betas controlling for all the other variables in the model. > To determine whether 24 is significantly different from 14, however, one would compute two models, as above, then compare the 2s of the different models to determine which lagged path significantly improves the goodness of fit or if both of them do. > If both lagged paths are significant, there is a bidirectional causal relationship. - There are several benefits for using SEM rather than simple regression equations: 1) it allows you to estimate all paths. - In Figure 4.5, for example, social support and PTSD symptoms are allowed to covary at Time 1 (the curved arrow). - Accounting for their covariations allows for more precise estimates for both the autocorrelations and the cross-lagged effects. 2) A second benefit has to do with unequal autocorrelations: not all variables are equally stable (or variable) across time. > For example, if one were to try to examine causal directionality between a personality trait and stress, the personality trait would be much more stable than the stress measure. > In other words, there would be less variance to be explained in Time 2 personality than in Time 2 stress, because the autocorrelations for personality traits would be much higher than those for stress. > Thus, the more stable variable always appears to be predicting the one that is more variable; however, in SEM, one can artificially ''adjust'' the autocorrelations (by constraining them to be more or less equal), which allows for a more accurate depiction of causality

Ipsative Change: Summary of Common Longitudinal Statistics

- Growth curve model: examines linear and nonlinear change over time. > Fixed effects refer to group-level effects, while random effects indicate individual differences in patterns of change. - Growth mixture model: Examines classes of change in trajectories over time

dose-response curve and dose-response effect: experimental design

- Health researchers often study the effects of a treatment on symptoms. - A major issue in question is what dose is adequate to significantly reduce symptom levels. - A dose-response curve plots symptoms against amount of medication: or it can also refer to the relationship between toxic substances and symptoms. - The classic finding is called a dose-response effect and refers to a linear relationship: the more of the toxic substance, the greater the impairment. > For example, Strawbridge, Wallhagen, Shema, and Kaplan (2000) found a dose-response effect between hearing impairment and problems in other domains: that is, the greater the degree of hearing impairment, the more the respondents reported psychosocial problems. - Sometimes, however, the relationship is not linear. > For example, there may be an optimum dose of a particular medication: lesser amounts may not have an effect, or greater amounts may not confer any additional benefits

Cluster Analyses

- In a related type of statistical procedure (factor analyses), cluster analysis, one groups individuals rather than variables. - For example, one could identify different clusters of individuals, one of whom may be high on depressive symptoms and the others low. - Conversely, one can use preidentified groups (e.g., married and unmarried) and then examine the differences between these groups on a number of items or scales. > This is similar to MANOVA but differs in that the items are used to predict membership in the groups. > If the groups truly are different, then the pattern of these items will correctly identify membership in the groups.

limitations of longitudinal designs: confounding aging and cohort effects

- In longitudinal studies excluding individuals with even one missing data point often result in losing substantial proportions of the sample, and researchers have developed innovative ways to estimate missing data (see Acock, 2005). - Longitudinal designs also confound aging effects with cohort effects, as one cohort may change in a way that other cohorts do not. > This is terribly important because nearly all of our information on aging has been gathered on the World War II cohort, and there is no guarantee that Baby Boomers or any subsequent cohorts will age in the same manner. > For example, WWII- and Vietnam-era veterans had very different war and home-coming experiences, and the effect of these differences on the aging experience is not as yet known. - Longitudinal designs also confound period (historical or time of measurement) effects. > This was first noted by Bradburn and Caplovitz (1965) who were conducting national studies on positive and negative affect. > This longitudinal study found a marked rise in negative affect, and a simplistic interpretation would be that negative affect increases with age. > However, they happened to take their second assessment at the time of the Cuban missile crisis when many thought that nuclear war was imminent, and the increase in negative affect reflected a period rather than an age effect. > Thus, longitudinal researchers sometimes use sequential designs to tease apart age, cohort, and period (historical or time of measurement) effects

regressions examining moderating effects

- In regressions examining moderating effects, one must first compute the interaction term, x1 x2. - In the hierarchical regression, x1 is entered in the first step, x2 on the second step, and the interaction term on the third. - If the interaction term is statistically significant, then one can conclude that smoking moderates the effects of hostility on blood pressure. - In other words, people who are high in hostility and who smoke may be much more likely to have high blood pressure than people who smoke but who are low in hostility. - In this instance, smoking enhances the effect of hostility on health. - If a moderator decreases the effect, then it is said to buffer the effect. > For example, problem-focused coping may buffer the effect of stress on health outcomes.

Sequenatial Analyses Summary: Parker and Aldwin

- In summary, in two of the three sequential analyses, Parker and Aldwin (1997) had significant age effects but no cohort or period effects. > They concluded that the increase in mastery in early adulthood is an age or developmental effect not restricted to a particular cohort or time period. > Values, however, were an entirely different story. Parker and Aldwin (1997) compared the primacy of family versus work values in the same sample and found massive period and cohort effects but no age effects. > Although most participants at all time points said that family was more important than work, the number of people who thought work was primary increased in the 1970s, especially among women, but decreased in the 1980s. > Thus, changes in personality appeared to be developmental and age related, but changes in personal values reflected the historical periods being studied.

Terminal drop: Longitudinal Designs

- In the 1960's, researchers discovered that if people are tested within about 2-3 years of their death they show some decline in cognitive functioning, especially on speeded tests. > Suppose you have a group of 200 people. You test themat age 60, 70, and 80. > When you test the group at 70, you notice that only 180 have returned. > You do some investigating and determine that of the 20 people who did not return, 17 died. If we compare the age 60 performance of those 17 people to the performance of those people who were alive at 70, we often find that those who died score lower than those who survived

Ipsative Change: statistics for assessing change

- Ipsative or idiothetic change (Lamiell, 1981) refers to the process of examining change at the individual level rather than the group level. - Lack of change at the group level does not necessarily mean that there is no change at the individual level. - an example of two individuals who changed in different ways: > Rachel had high scores on introversion at Time 1 but was low at Time 2, and Jill showed the opposite pattern. > The difference between Time 1 and Time 2 for this group of two individuals would be 0—the group average did not change, but the individuals showed different and opposite trajectories. - One of the simplest ways to examine ipsative change is simply to look at turnover rates, or whether a person stays in the same category over time. > For example, using the Normative Aging Study (NAS) alcohol data archive, Levenson, Aldwin, and Spiro (1998) showed dramatic differences between group and individual stability in problem drinking. > At the group level, the percentage of NAS men who reported problem drinking remained roughly stable at about 5% at all three time points (1973, 1982, and 1991). > However, there was a very high turnover rate: almost none of the men who were problem drinkers in 1973 still had problems in 1982, and most of the problem drinkers in 1982 did not report problem drinking in 1991. In fact, out of 1,056 men who had data at all three time points, only 25 (.02%) reported problem drinking at all three time points. So, although the group rates of problem drinking - Ipsative or idiothetic change (Lamiell, 1981) refers So, although the group rates of problem drinking appeared stable, the ipsative data showed that individuals were not at all stable. Thus, one cannot generalize from group-level data to individuals. p 74 fig

reconciling different findings

- It is not unusual for some studies to find positive effects, others to find negative effects, and yet others to find no effects. > For example, early studies found that older adults had high rates of depression, but these were largely based on clinical or help-seeking samples. > Epidemiological studies that assessed random samples sometimes found that younger adults were more likely to be depressed than older adults; others found no relationship between age and depression (see Aldwin, 1991). - when faced with contradictory evidence, it is important to understand that the process of reconciling disparate findings can often provide a much more detailed and precise understanding of the phenomenon. > If there are differences between studies, these must be due to the types of samples used, the types of measures assessed, and/or the types of analyses conducted. - In the problem just cited about aging and depression, all three types of problems were relevant. > The results varied by the age of sample: symptoms decreased from young adulthood to midlife, then rose in late life; therefore, studies that sampled primarily young adults and the middle-aged tended to find a negative relationship between age and depressive symptoms, whereas those whose samples were older often found a positive one. - Type of measure used was also important. > Studies using measures of depression that included physical symptoms such as fatigue, sleep problems, or aches and pains were more likely to find that older adults were more depressed—not because they were unhappier but because their health was poorer. > Studies that focused just on negative affect often found younger adults were more likely to be depressed. - The type of analysis used was also critical. Most simple statistics examining the association between two or more variables assume that the association is linear (which is why regression is called ''linear regression''); however, many associations are nonlinear, and special procedures are needed to identify nonlinear relationships (see Aldwin et al., 2001). - In this case, there was a nonlinear relationship, a J-shaped curve, between age and depressive symptoms.

Analyzing Multiple-Point Longitudinal Data: STATISTICS THAT PREDICT CHANGE

- Longitudinal studies are becoming more sophisticated, and often they have more than two data points. - Although one could use a series of correlations, regular MANOVAs, or SEMs to analyze data occurring at multiple points in time, missing data often results in unbalanced data sets, with different people having different numbers of valid follow-up assessments, making traditional analyses difficult or impossible to interpret. MANOVAs, for example, typically require complete data at all time points. If different people are in the sample at the different time points, the change in mean could be due to sample composition rather than time. - A number of different ways of examining multiple-point data have emerged in the past few years - These types of analyses often examine change over multiple time points, do not require equal time periods between assessments, and also do not require every individual to have data at each assessment, as they can estimate missing data: 1) The simplest way of analyzing this type of data is to compute individual trajectories by fitting regression lines for each individual. > This yields a predicted intercept (baseline) and slope (B) for each person. > In a two-stage growth modeling approach, the resulting intercepts and Bs can be used as dependent variables in a second equation to predict change. > Aldwin, Leven- son, Spiro, and Bosse ́ (1989), again using NAS data, computed individual trajectories, plotting symptoms against time, and showed a clear linear increase in physical symp- toms over time; on average, men in the NAS sample developed one new symptom every 3 years. > As Affleck et al. (1999) pointed out, however, there are a number of problems with simply computing individual regression lines: they do not control for the correlated errors common in longitudinal data sets, they may unduly capitalize on chance, and the measure of the goodness of fit of the line, root mean-squared error (RMSE), may be inaccurate on an individual basis; RMSE measures how much the individual data points deviate from the predicted line; a large RMSE indicates that the fit is not very good, a small one that the fit is good; however, if someone is stable and has an absolutely flat line, there is no (0) deviation, and the RMSE cannot be assessed accurately, as it would involve dividing by zero. 2) Generalized estimating equations, or GEEs are one way of examining trajectories over time and provide the advantage of being able to use time-varying covariates. > In other words, GEEs can answer questions such as whether individuals' curves differ as a function of changes in some type of variable. > For example, Schnurr, Spiro, Aldwin, and Stukel (1998) examined whether individuals' health trajectories changed as a function of exposure to trauma; they found that individuals who were exposed to both combat and noncombat trauma were likely to have steeper curves than those who were exposed to only one type of trauma (or who were lucky enough to avoid all trauma); exposure to both combat and noncombat trauma resulted in an increased rate of symptoms change with age. > The problem with GEE designs is that the metric is a little awkward: to examine symptom change, for example, a two-way interaction term—Symptoms Time—is computed; to examine one covariate, a three-way interaction term is computed—Symptoms Time Trauma; obviously, the more variables one has, the more complex the interaction terms, and power to examine multiple-level interactions can often be insufficient. 3) A more sophisticated way of estimating individual trajectories is to compute a growth curve model also called a hierarchical linear model > In HLM models, the average trajectory is computed first (fixed effects), and then individual trajectories are computed as variants of the average curve, or how they differ (random effects). > This yields an estimated curve for each individual to avoid some of the problems with correlated errors because these error terms can be modeled and the curves adjusted accordingly. > For example, Aldwin, Spiro, Levenson, and Cupertino (2001) examined changes in psychological and physical symptoms over time using random effects models; to examine nonlinear patterns of change over age, they included polynomials in their equations (age2 and age3); using cluster analysis (similar to factor analysis, but grouping people rather than variables), they then identified patterns of change over time; as Figure 4.6 shows, the average slopes, or fixed effects, were rather simple: physical symptoms increased linearly, whereas there was almost no change in psychological symptoms; figure 4.7 shows the random effects models for physical symptoms: some individuals showed little or no change in the number of symptoms reported, but others showed exponential change. Individuals in particularly poor health showed asymptotic curves, whereas others resembled sine curves; Averaging the slopes ignored the astounding amount of individual differences in health over time. 4) A new type of analysis, growth mixture models (GMM) calculates trajectories and groups them in a much more sophisticated way > Aldwin et al. (2011) used GMM to examine patterns in stressful life events over time. > As mentioned earlier, the purpose of GMM is to examine classes of trajectories by examining increasingly more complex models and examining goodness-of-fit indicators. > For example, we began by fitting a model specifying one class then calculated a second model specifying two classes; if the good- ness-of-fit statistic decreases significantly, then the two-class statistic is thought to be an improvement; one repeats this process until either the fit statistic starts increasing, indicating that the model is becoming worse, or until the groups are too small to be meaningfully interpreted. > Using this technique on stressful-life-event data spanning 18 years, we found that there were four types of classes of trajectories, three of which showed life events decreasing with age but one of which showed a nonlinear increase (see Figure 5) Latent growth curve models are another way of examining multiple-point designs. > This model is a variant of an SEM approach and can handle the type of missing data often found in multiple-point longitudinal designs as well as correlated errors. > It also can provide a more direct means of predicting change. > A drawback, however, is that latent growth curves with more than three or four points tend to be difficult to analyze, especially if one starts examining correlated errors. (Take Figure 4.3 and multiply it by 5, for five time points, then start drawing arrows between all the different times, to generate a truly complex figure.) > In standard latent growth curve models, the goal is to determine which variables predict change rather than showing patterns in how individuals change; thus, it is more a variable-centered than a person-centered approach. > Growth curve analyses are still relatively new, and there are a number of problems to consider for which there are currently no set answers. > For example, in examining developmental change, should one use age or time as the dependent variable? If one enters polynomial terms to examine nonlinear change with age, the age terms should be centered to control for multicollinearity, but if random effects models are used, should the mean age be used as the center or the individual's age? In latent growth curves, what criteria should be used to determine possible causal directionality? In GEEs, does one omit the outliers to smooth out the curve or does omitting the outliers lose the rare (and possibly most interesting) cases? In estimating missing data, should the population average or the individual average be used? 6) One of the fastest growing areas is multiple imputation analyses in which the other variables in the data set are used to estimate or impute what the individual's answer would have been. > This allows for a more complete data set but requires assumptions about missing data that are seldom met. > For example, technically, missing data should be ''missing at random'' (or MAR); in other words, missing data should be randomly missing, not due to any external or design factor, an assumption which is often difficult to meet. > Bayesian statistics are becoming more common for the imputation of missing data, as they may not have as restrictive assumptions p 80 fig

Correlational Design

- Many studies do not divide people into groups but instead simply examine the associations among groups of variables. - To assign variables to independent or dependent status, one must have a strong theoretical rationale. > For example, when studying the relationship among personality, coping, and symptoms, it may be very difficult to assign causal directionality (i.e., which variables cause the others). > A reasonable hypothesis is that individuals who are high in hostility cope with problems by ex- pressing negative affect (e.g., yelling), which in turn raises their blood pressure. It is entirely possible, however, that highly variable blood pressure may lead to overreacting in stressful situations, resulting in hostile behaviors and outlooks (Kasl, 1983). - Thus, causal directionality cannot be determined with cross-sectional, correlational data, but one can find interesting relations that can be further investigated.

age-related designs

- Many studies of the effects of aging use quasi-experimental designs because it is not possible to randomly assign people to different ages. > However, some studies examine age effects by sampling young and old subjects, randomly assigning them to experimental conditions, and then testing for interactions with age. - The statistics involved with this type of design are fairly simple, usually involving t-tests or ANOVAs (see Table 4.1). - Other studies simply use natural groupings of different ages or follow individuals over time and thus use quasi-experimental designs that often have more complex statistics. > Therefore, this section focuses on quasi-experimental designs and the statistics required to analyze them.

mediating effects

- More likely (than a direct effect) there is an association between two variables, but the effect is actually mediated by a third variable. - For example, hostility may be associated with higher blood pressure levels because people who are high in hostility are more likely to smoke, which in turn is associated with higher blood pressure. > This is called a mediated or indirect effect.

Multivariate analyses of variance (MANOVA)

- Multivariate analyses of variance (MANOVAs) are used when there are multiple dependent variables that are intercorrelated. - MANOVAs can control for correlated errors that reflect systematic biases in response patterns. > For example, some people fill out questionnaires very quickly, and, when confronted with a scale of several items rated from 1 to 5, they may just circle all the 5s. > If they do this with several scales, high correlations among the scales may result, but these are probably spurious (hence the term, correlated errors). - Thus, if a researcher has a set of dependent variables that are intercorrelated, it is much better to calculate a MANOVA that yields a multivariate F rather than calculating several different ANOVAs.

Timing: Special Design Issues in Health and Aging Research

- Must distinguish between a proximal, or immediate cause and a distal, or distant, one - Examining the relationship between assessing biomedical outcomes - Establish what is the response latency and recovery times are for any given biomarker

Two-Point Designs: STATISTICS THAT PREDICT CHANGE

- Often researchers want to use baseline data (Time 1) to predict some outcome (Time 2). - The simplest way of doing this is to use either multiple regression if the dependent variable (y) is continuous or logistic regression if it is categorical; however, two- point designs can quickly become very complicated. - In biomedical and psychosocial research, the outcome variable may occur over a period of time. > For example, analyses predicting outcomes such as death or the recurrence of cancer will typically follow the sample over a period of several months or years. > Not only will there be individual differences in whether the outcome variable occurs but there may be differences in the timing of that occurrence. - To analyze this type of data, researchers typically use a special form of logistic regression called Cox proportional hazards models (Cox & Oakes, 1984). > In these models, occurrences of the outcome are referred to as events (e.g., mortality); cases that do not have the event by the study's end are censored: that is, they are still in the study but counted as having survived for at least the study's duration. > Relative risk ratios (RRs) are calculated that indicate the risk by period of time (usually years). > An RR of 1.7 refers to a 70% increase in the risk of the outcome per unit increase of the independent variable relative to a comparison group, whereas > .7 refers to a 30% decrease in such a risk. - Epidemiologists often categorize the independent or predictor variable, in which case one of the categories is used as the reference group. > For example, it is easier to comprehend a finding that pack-a-day smokers have a 70% increased risk of mortality compared to nonsmokers than a finding that there is an increased risk of 3.5% per every cigarette smoked per day. - Sometimes, however, researchers want to predict change in an outcome variable. > If you have a continuous dependent variable, the easiest way to do this is simply to compute change scores (y2 - y1) as the dependent variable; however, for reasons outlined earlier, change scores fell out of favor in psychological research, and residualized regressions became the most common way of examining predictors of change. - Residualized regressions are hierarchical regressions in which the Time 1 dependent variable is entered first and thus covaried out, and other variables are used to predict the adjusted Time 2 scores. - For reasons that are not clearly understood, some researchers have begun using change scores as dependent variables in a residualized regression equation. > This makes very little statistical sense because it results in the same variable being on both sides of the equation. y2 y1 =y1 +b1x1 +b2x2 +a Solving the equation yields: y2 =2y1 +b1x1 +b2x2 +a > Logically, there is no reason why twice the level of the dependent variable at Time 1 should be used to predict the dependent variable at Time 2, so this may result in invalid negative betas. > Thus, one should either use change scores as the dependent variable or residualized regressions, but not both.

Observational Designs

- Often used at the very initial stages of a research project when there is a phenomena of interest and the researcher simply wishes to understand a particular culture, situation, or problem better - Researchers in the social sciences will often conduct ethnographies or other qualitative studies when either trying to formulate a question for a newly observed problem, or if they believe it is to soon to develop questionnaires - Qualitative studies may be highly useful when there are contradictions or anomalies in more quantitative studies and a more in-depth perspective is needed

Random Assignment & Placebo Effect: Experimental Designs

- Random assignment to groups is absolutely critical to determine whether or not a given manipulation or treatment works. - Especially in clinical trials that examine the efficacy of a drug, these should be double-blind studies in which neither the medication administrator nor the research participant knows whether or not that person has been assigned to the experimental group. > Instead, an elaborate system of codes is used to track which participants receive the drug and which receive the placebo. > At the end of the study, these codes are broken and researchers can analyze the data. - This controls for the placebo effect: the tendency for people to positively evaluate any treatment, even though there may be no physiological effects attributable to the treatment.

Meta-Analyses: Special Design Issues in Health and Aging Research

- Researchers identify as many studies as possible that have analyzed similar patterns of development and aging. - This is a heavily used but very limited technique. - It was developed in the context of medical research on simple straightforward questions. Is smoking related to cancer? > At the time the decision was made to place Warning Labels on cigarette packs there were dozen of studies using many different methods. > Meta-analysis was developed to synthesize the data across studies. - A powerful way of summarizing the literature and developing hypotheses to explain contradictory findings, which can lead to definitive studies to resolve contradictions in the literature. - It's power is limited by the limits on meta-analysis. > We may have many studies that use SEM to model smoking data, but meta-analysis cannot evaluate SEM models. > Investigators have to break the model into smaller pieces but they cannot capture the who thing.

Studying age-related differences vs. changes: cross-sectional and longitudinal designs

- Several decades ago, researchers recognized that studying age differences is very different from studying age-related change. - Age differences are studied using cross- sectional designs, which compare different age groups at one point in time. - Age- related changes are examined using longitudinal designs, which follow people over time. - The two designs can yield varying results. - Finding age differences in a phenomenon between different age groups does not mean that there are age-related changes.

Factor analyses

- Sometimes, a researcher is interested in identifying items that group together, like trying to find subscales in a depression inventory. - Factor analyses are typically used in this type of inquiry. - There are two basic types of factor analyses: 1) principle components seeks to identify the items that most closely track a construct 2) principle axes seeks to identify different dimensions within a larger construct. > ex: say a researcher was constructing a scale for depressive symptoms and had generated 100 items. > If the idea was to identify the 10 items that best indexed depressive symptoms, then the researcher would conduct a principle components analysis. > If, however, he thought that there were several dimensions to, or types of, depressive symptoms, then he would do a principle-axes analyses to identify different subscales. - Note that factor analyses have different types of rotations, which allow the factors to be more or less orthogonal (unrelated) and have different ways of allotting variance across factors. - Thus, it makes little sense to do a principle-components analysis with varimax (orthogonal) rotation, as the latter seems to maximize variance across different dimensions, which the former seeks to put most of the variance on the first dimension.

bivariate vs multivariate statistics

- Statistics can be grouped depending on whether they are bivariate (examining only two variables, such as t-tests, correlations, or 2s) or multivariate (examining more than one variable, such as multiple or logistic regression).

Test-retest (Practice) Effects: Longitudinal Designs

- Suppose you are doing a longitudinal study of memory. > Your study involves as king participants to view 20 lists of 20 words. > They are tested at age 75, 77, and 79. > Some participants think they could do better, and without your knowledge they start practicing with lists of words, some even buy a book on memory techniques that do have a positive effect if you practice their methods. - In this study there is a control group that only received the test at age 77, and another that is tested only at age 79. > These control groups allow us to see whether there is a test-retest effect which improves the average performance of the 77 year-olds that had been test at age 75. > Some people use the term practice effect but there are any number of ways performance could be improved.

Table 4.1 summary of commonly used cross-sectional statistics and their uses - p 63

- Table 4.1 presents a summary of commonly used cross-sectional statistics and their uses, divided by whether they test differences or associations. - For example, t-tests assess mean differences between two groups, whereas correlations examine whether two variables covary using a standardized metric (1 to +1). - A positive correlation between two variables means that people who are high on one measure also tend to be higher on the other one (e.g., education and income), whereas a negative correlation means that being high on one variable is associated with being low on the other.

differences between mediating and moderating effects

- The difference between mediated and moderated is often very subtle and the two are easily confused. - Further, the same variable may be both a mediator and a moderator. - In some designs, mediators may also be considered a type of confounder: > For example, take the association between hostility and heart disease. Why do people who are high in hostility have higher heart disease rates? > As mentioned earlier, this relationship may be due to a third variable such as smoking. > The simple way to test for this is to do a hierarchical (or ordered) regression equation in which one enters hostility in the first step and smoking in the second step. > If the income variable significantly reduces or eliminates the relationship between hostility and heart disease, then one can say that there is no direct relationship between ethnicity and crime: that everything is mediated through (or confounded with) smoking. > A researcher could also ask whether smoking moderates the relationship between hostility and heart disease. > One would compute an interaction term between smoking and hostility. > It is possible that individuals who smoke and who are high in hostility have higher levels of heart disease; thus, smoking could be both a mediator between and a moderator of hostility and heart disease. - The distinction between mediated and moderated is subtle but perhaps can best be remembered by the differences between the associated statistical analyses. - Both designs typically use hierarchical multiple regression, which enters variables in the equation in a specific order or step. - To examine mediating relationships, x1 is entered in the first step, so the direct relationship between x1 and y is examined. > In the second step, the hypothesized mediator, x2, is added. > If the between x1 and y is significantly decreased, then one can conclude that the effect of x1 on y is through x2. > In the example above, smoking (x2) mediates the relationship between hostility (x1) on blood pressure (y).

Observational Studies - Qualitative Studies

- The fourth type of design is purely observational. - These are often used at the very initial stages of a research project when there is simply a phenomenon of interest and the researcher simply wishes to understand a particular culture, situation, or problem better. - Researchers in the social sciences will often conduct ethnographies or other qualitative studies when one is either trying to formulate a question for a newly observed problem, or if it is unlikely that the question could be answered through questionnaires. > For example, in the early days of the AIDS epidemic in San Francisco, public health officials only knew that gay men were dying, but not why. > Some of these officials conducted qualitative studies, living in the Tenderloin (a rough district frequented by male and female prostitutes and drug users) in order to establish relationships with individuals to try to determine what factors put these individuals at risk. - Qualitative studies may also be highly useful when there are contradictions or anomalies in more quantitative studies and a more in-depth perspective is needed.

The study of age-related changes in health, cognition, and psychosocial functioning demands new ways of designing studies and analyzing data

- The past decade has witnessed the development of new techniques that allow researchers to address more interesting and complex questions - As research addresses more complex questions and we become aware of shortcomings in existing methods, new methods evolve. - Even though you may not be doing research, in order to read and understand research articles you need to have some familiarity with the terms and the decision making of researchers. - These methods and statistics often come with their own terminology which makes it difficult to understand

Ipsative Change (Idiothetic Change)

- The process of examining change at the individual level rather than the group level - One of the simplest ways to examine ipsative change is simply to look at turnover rates, or whether a person stays is the same category over time - One cannot generalize from group-level data to individuals - A second way of examining change at the individual level is to calculate change or difference scores

Meta-Analyses: SPECIAL DESIGN ISSUES IN HEALTH AND AGING RESEARCH

- The process of trying to reconcile findings across studies is called a meta-analysis (see Card, 2012). - In meta-analyses, researchers identify as many studies as possible that look at a given phenomenon and try to account for the differences. - Sometimes, there are relatively few studies, and these have used very disparate measures, so the best one can do is to determine what proportion of the studies found the relationship in question. > For example, Tennant and McLean (2001) reviewed the literature examining whether people who were recently widowed were more likely to die in the following year; they reported that 9 of 14 studies showed a positive association. - Sometimes, there are many studies, and one can statistically examine effect size depending on the type of sample or the type of measure used. > For example, Miller, Smith, Turner, Guijarro, and Hallet (1996) conducted a meta-analysis of the relationship between hostility and health. > They concluded that there was a general relationship between hostility and coronary heart disease but that the magnitude of the effect depended in part on the particular hostility measure used as well as the type of study design. > The structured interview provided stronger and more consistent results than self-report measures. > Furthermore, studies using case control designs that contrasted sick with healthy individuals provided the strongest results, whereas those that restricted the range of illness outcomes (e.g., those that examined only patients with coronary heart disease) showed the weakest results. - Meta-analyses are a powerful way of summarizing the literature and developing hypotheses to explain contradictory findings, which can lead to definitive studies to resolve contradictions in the literature.

Analyzing Multiple-Point Longitudinal Data: Statistics That Evaluate Change

- The simplest way is to compute individual trajectories by fitting regression lines for each individual - Two-stage growth modeling approach, the resulting intercepts and Bs can be used as dependent variables in a second equation to predict change - Growth curve model is a more sophisticated way of estimating individual trajectories - Hierarchal linear model or multi-level models are used in models that are nested - Growth mixture model calculates trajectories and groups them in a more sophisticated way

Quasi-Experimental Designs: powerade example

- Used when it is neither practical nor ethical to randomly assign people to experimental groups. - We might study people who drink powerade during and following exercise and compare them to people who do not. > These people are voluntarily doing an experiment on themselves! - Is critical to statistically control for other differences between the groups. - In this case we might to enroll equal numbers of men and women in our study, keep track of their body mass index, and some measure of fitness because powerade might work differently on fit and non fit people. - Causal language, as in "drinking powerade causes rapid recovery following exercise," should be avoided and instead talk about associations. Drinkers of powerade appear to recover more quickly but we have not ruled out all competing explanations.

Timing: SPECIAL DESIGN ISSUES IN HEALTH AND AGING RESEARCH

- There are three major issues concerning the timing of variables in health and aging research: 1) one must distinguish between a proximal, or immediate, cause and a distal, or distant, one. > For example, the proximal cause of having a heart attack may be eating a holiday meal loaded with fat or shoveling snow; the distal cause may be having been a low-birth-weight baby, whose liver never developed properly, resulting in chronic high cholesterol levels (Barker, 1999). > When examining distal causes, it is useful to construct a plausible chain of events that links distal and proximate causes. > Of course, having longitudinal data is very helpful when linking distal and proximate causes. 2) occurs when examining the relationship between assessing biomedical outcomes. > Many of the biomarker outcomes that psychologists are interested in demonstrate daily fluctuations called diurnal rhythms. > For example, the stress hormone cortisol typically peaks shortly after rising and then decreases throughout the day; thus, any study that examines these outcomes must take into account time-of-day effects; an even better approach is to examine three to four assessments of cortisol each day to determine the characteristics of the diurnal rhythm, that is, not only the peak amounts but also the rate of change in the slope 3) it must be understood that different biomarkers have different stimulus latencies and recovery periods. > For example, in response to a stressor, epinephrine increases very quickly within seconds, but cortisol may take 20 minutes or so to increase; however, epinephrine also ''clears'' (or dissipates) far more rapidly than cortisol; this can create many Type 2 (false negative) errors if this is not taken into account. > For example, say one wanted to understand differences in biomarker responses to stress using an experimental paradigm with mild pain as a stimulus; if the researchers wait 5 minutes to collect the blood sample, chances are likely that they will find absolutely nothing; the change in epinephrine will have already occurred, but the cortisol reaction won't yet have occurred; thus, it is important to establish what the response latency and recovery times are for any given biomarker. > Further, there are often individual differences in baseline levels of biomarkers which may be due to a number of variables, including genetic differences, nutritional status, health status, and so on. > In order to adequately assess the effects of stress, typically researchers examine change from baseline and then time to recovery; of course, this is generally done in a laboratory setting; however, there are new techniques for ambulatory assessment of biomarkers which are being developed which might open exciting possibilities for new research.

sequential designs: age-related designs

- There are three types of sequential designs: cohort-, cross-, and time-sequential (Schaie, 1977). - Remember that age refers to a person's chronological age, cohort to the year in which he or she was born, and period to the time of measurement. - The differences among these three designs are depicted in Table 4.2, which is based on data from a study of change in mastery from young adulthood to midlife by Parker and Aldwin (1997). - This study used data from the Davis Longitudinal Study, which follows several cohorts of college alumni (the classes of 1969, 1979, and 1989).

Cross-Sequential Designs: age-related designs

- These designs examine cohort and period effects but confound age (see panel B, Table 4.2). - The cross-sequential design shown in Table 4.2 examines two cohorts (1947 and 1957) at two periods (1979 and 1989), but it includes measurements at three different ages: 22, 32, and 42. - Again, cross-sequential designs confound age; here, there are two assessments of 32-year-olds but only one assessment of 22-year- olds and one of 42-year-olds. - Parker and Aldwin (1997) examined whether the two cohorts changed in similar ways in mastery from 1979 to 1989 and found that they did not. > The older cohort did not change, but the younger cohort increased in mastery. > Therefore, this change was probably not a period effect because the two cohorts did not change in the same way. > But to confirm this, Parker and Aldwin also performed a time-sequential design.

Time-Sequential Designs: age-related designs

- Time-sequential designs examine two different age groups at two different time periods. - To do this, however, requires three cohorts. In these designs, only one cohort is studied longitudinally (the 1957 cohort in Table 4.2), but it is compared at two different periods with a different cohort (hence three cohorts). - In other words, in the time-sequential design illustrated in Table 4.2, panel C, the 1957 cohort is followed from age 22 to age 32. In 1979, this cohort (aged 22) is compared to the 1947 cohort (aged 32). In 1989, the 1957 cohort is now aged 32, and it is compared to the 1967 cohort (aged 22). > This design answers the question, ''Is the difference between 22- and 32-year-olds in the year 1979 similar to that between 22- and 32-year-olds in 1989?'' > Thus, time-sequential designs have two age groups and two time periods but require three cohorts, confounding cohort. - When Parker and Aldwin (1997) performed this time-sequential analysis, they found nearly identical differences in mastery between 22- and 32-year-olds at both time points, which suggested there was indeed an age or developmental change in mastery between the 20s and the 30

Summary of Common Cross-Sectional Statistics p 63 - Parametric (continuous) Bivariate (2 variables)

- Type of data: > Parametric (continuous) Bivariate (2 variables) - Are There Differences: > t-test: Whether two groups differ on a continuous dependent variable y - Are There Associations: > Correlations (r): Significance of the association between two continuous variables > Canonical Correlations: Significance of the association between two sets of variables

Summary of Common Cross-Sectional Statistics p 63 - Cluster Analysis

- Type of data: > Cluster Analysis: Identifies types of people that are similar - Are There Differences: > Factor Analysis: Identifies the patterns of associations of items within a construct - Are There Associations:

Summary of Common Cross-Sectional Statistics p 63 - Discriminant Analysis

- Type of data: > Discriminant Analysis: Identifies the patterns of differences between 2+ groups on a set of variables? - Are There Differences: > Structural Equation Modeling (SEM): Similar to path analysis, but can include latent variables (composed of factor- weighted items - Are There Associations:

Summary of Common Cross-Sectional Statistics p 63 Multivariate (2+ variables)

- Type of data: > Multivariate (2+ variables) - Are There Differences: > Analysis of Variance (ANOVA; F)—Whether 2+ groups differ on y - Are There Associations: > Multiple Regression: Identifies the best predictors of y

Summary of Common Cross-Sectional Statistics p 63 - Multivariate (2+ variables)

- Type of data: > Multivariate (2+ variables) - Are There Differences: > Loglinear: Whether different groups have different distributions on multiple categorical variables - Are There Associations: > Logistic Regressions (odds ratios): Identifies best predictors of a dichotomous outcome (e.g., death)

Summary of Common Cross-Sectional Statistics p 63 Multivariate analysis of variance (MANOVA)

- Type of data: Multivariate analysis of variance (MANOVA: Whether 2+ groups differ on a set of variables? - Are There Differences: > Path Analysis: Identifies the pattern of associations across multiple variables - Are There Associations:

Summary of Common Cross-Sectional Statistics p 63 - Nonparametric (categorical) > Bivariate (2 variables)

- Type of data: Nonparametric (categorical) > Bivariate (2 variables) - Are There Differences: > Chi-square (2): Whether the observed categorical distribution differs from the expected one - Are There Associations: > Eta: Significance of the association between two categorical variables

Quasi-Experimental Designs

- Useful when it is neither practical nor ethical to randomly assign people to experimental groups. > For example, in studying the long-term health effects of parental divorce in childhood, one cannot randomly assign children to groups (the parents cannot be asked to divorce just so one can study its effect on their children) > Thus, the researcher must employ a quasi-experimental, or cohort, design in which one identifies people whose parents divorced when they were children and compare them to similar people whose parents did not divorce. > The hitch here is ''similar': children of divorce may differ in other ways, such as exposure to an abusive or mentally ill parent or have greater financial difficulties before or after divorce. - Thus, it is critical to statistically control for other differences between the groups to identify those effects that could be the result of divorce. - Researchers who use quasi-experimental designs are generally very careful to avoid using causal language (e.g., divorce causes behavioral problems in children) and instead talk about associations (e.g., parental divorce is associated with behavioral problems).

when to use which statistic

- Which statistic one uses depends on the question, the research design, and the type of data that are available. - a good question literally dictates the type of design and the statistics used in a study. - Age-related designs and statistics are even more complicated but allow one to address more complex questions.

Statistics

- answers two basic questions: (a) Is there a difference between two (or more) groups on a given variable (b) is there an association between two or more variables? - These two questions are often simply restatements or mirror images of each other. -The question, ''Are low-birth-weight babies more likely to have health problems later in life than normal-birth-weight babies?'' compares two groups. > This question is identical to ''What is the association between birth weight and subsequent health problems?'' - Statistics are used to establish whether the observed effects are ''real'' (significant) or due to chance.

Multivariate analyses of variance (MANOVAs)

- are used when there are multiple dependent variables that are intercorrelated. - Aldwin makes frequent reference to this design. > These are very careful in aging research. > For example, an investigator may measure four types of memory function at ages 40 60 and 80. > This is a quasi experimental design because we cannot assign people to age (Zap you are 40 years old!) but instead of one variable such as IQ, there are four variables (or variates). > This makes it a multivariate study, and we use MANOVA to evaluate the outcome.

Regression equation: Common Statistical Techniques

- basically a series of partial correlations. - Does much the same thing but with several variables so that one can examine the contribution of each variable dependent of the others in the equation

Research designs: purpose

- basically frameworks for answering research questions. - The research question drives the selection of the sample, the measures, the assessment framework, and the analyses. - A simple question is: what is the association between x (the independent or predictor variable) and y (the dependent or predicted variable)? > This very simple question should only be asked in exploratory research. - More sophisticated questions address problems in the research literature and specify explicit hypotheses that contrast different theoretical models underlying previous research. > For example, a researcher might notice that one group of studies showed that age is positively associated with depression, but another group of studies showed that age is negatively associated with depressive symptoms. > Examining these studies closely, she might notice that the first group used casesness, or diagnoses of depression, while the second group simply assessed the frequency of depressive symptoms. > Alternatively, it might be that the two sets of studies had samples with different age ranges, with the first group including more individuals who are very old and the second having a more restricted age range. > Thus, she could design a study that includes both clinician diagnoses and self-reported depressive symptoms to see if the same associations occur in the same group of people who range in age from very young to very old. - The question thus dictates the choice of sample, the measures and, as we shall see, the analyses.

Comparison of the Different Designs

- different designs may result in very different findings. - Clinical studies, especially those using case control designs, may overestimate the effect size (or magnitude of the effect). - Large surveys using quasi-experimental designs that gather more random samples often underestimate the effect size, because individuals who are ill or troubled tend not to respond. - Experimental studies may not find an effect because their sample sizes are too small, or because they use too little of the treatment, assess the outcomes at inappropriate times, and so on. - Correlational studies often use convenience samples (e.g., college students) whose generalizability may be limited and certainly cannot address causality. - Qualitative or observational designs may also have limited generalizability. > Nonetheless, all of these types of designs can yield valuable insights from their different perspectives on problems depending on how the data are analyzed.

Multivariate: Common Statistical Techniques

- examines several variables on one sample. - Analyzed using MANOVA or multiple regression)

most common research designs

- experimental - quasi-experimental - correlational - observational

mean-level change: statistics for assessing change

- figure 4.1 illustrates an example of mean-level change. > This figure depicts a personality trait, introversion, which has a normal distribution and a mean of 5 at Time 1. > At Time 2, the personality trait still has a normal distribution, but the entire distribution has shifted two points to a mean of 7. > This indicates that, as a whole, the population increased in this personality trait with age. - The simplest way to assess whether there is a change in mean level over time is a paired t-test. - A standard t-test compares two different groups (a between-subjects analysis). > If individuals are followed over time, this is a within-subjects analysis. > Thus, a paired t-test can determine whether two groups increased or decreased on a variable over time, but it cannot tell you very much about why they changed because it does not include any way of assessing predictors of the change. - A better way of assessing change is a repeated-measures MANOVA. - A repeated-measures MANOVA also assesses mean-level change over time (within- subjects analysis), but it also allows examination of whether or not different groups are changing similarly (between-subjects analysis). - The MANOVA yields F statistics for time as well as for group. > The statistical significance of the interaction term between the grouping and time variables indicates whether the two groups differ in how they have changed over time. - For example, a repeated-measures MANOVA can determine whether men and women show the same pattern of change or whether they show different patterns over time. > We could hypothesize that personality changes with age differently in men and women. > Men might become less extraverted and women might become more extraverted. > Thus, we would hypothesize a Gender Time interaction. - Repeated-measures MANOVAs are used to analyze both cohort- and cross- sequential designs. > In cohort-sequential designs, the independent variables are age and cohort, and in cross-sequential designs, the independent variables are cohort and period. - To use a repeated-measures MANOVA, one must have a balanced design; that is, all respondents must have assessments at all points in time. - However, simple ANOVAs must be used to analyze time-sequential designs, as only one cohort has two data points; the other two cohorts have only one data point each, and thus between-subjects statistics cannot be used. > For example, in Table 4.2, only the 1957 cohort has two time points; the other cohorts have only one. > Therefore, a repeated- measures MANOVA could not be used, because only one group actually was assessed twice. p. 70 fig

Statistics for Assessing Change

1) Absolute or Mean level change refers to changes in mean level over time 2) Relative change refers to shifts in rank order 3) Structural change refers to where or not the factor structure is similar across age or time 4) Idiothetic or ipsative change refers to changes in individuals. - This is the least commonly used type of change but it can be very interesting. - Suppose I test people every 5 years for 25 years. > I look at the pattern of change for each person. > I discover that there are people who never change, some who score higher at each test, some who show a drop after 10 years and never improve after that. > Researchers have found patterns like these in several studies. > They they try to find characteristics that distinguish the different groups.

Sequential Designs

1) Cohort-sequential designs - Follow two or more cohorts over the same ages but confound time 2) Cross-sequential designs - Examine cohort and period effects but confound age 3) Time-sequential designs - Examine two different age groups at two different time periods - Requires three cohorts - Only one cohort is studied longitudinally but it is compared at two different periods with a different cohort

SUMMARY

The past decade has seen the emergence of new ways to analyze longitudinal data. These analytical techniques can answer more sophisticated questions, and they pro- vide the opportunity to examine complex designs such as those needed to examine multivariate or transactional models. However, a serious problem with these types of analyses is that they are all extremely complicated and require sophisticated un- derstanding. Interpreting the output of such programs is often more of an art than a science. Nonetheless, mastering them can help answer some fascinating research questions, including what predicts different rates of aging.

Bivariate: Common Statistical Techniques

examines only two variables such as t-tests, correlations, or X2s

Categorical variables

refer to groups that cannot be added together, such as where someone was born or type of illness.

Experimental Designs: procedure

research participants are randomly assigned to an experimental group that receives the treatment or manipulation and a control group that does not (or that receives a placebo of some sort).

Structural equation modeling (SEM): Common Statistical Techniques

simultaneously estimates the relationships among several variables

Nonparametric statistics

such as chi-square (X2) analysis are used if both the independent (predictor) and dependent (outcome) variables are categorical.

Placebo effect: Experimental Designs

the tendency for people to positively evaluate any treatment, even though there may be no physiological effects attributable to the treatment

Continuous variables

those that form a series, such as age, number of words remembered, hours slept, or number of symptom

Partial correlation: Common Statistical Techniques

— controls for one variable a when correlating two other variables (b with c). > For example, I do a study of the effects of alcohol on a maze solving task. > There are 5 different doses of alcohol, so I want to correlate dose with performance. But are these results influenced by body mass index? - Partial correlation allows us to "partial out" the influence of BMI .