Cohort Analysis of the Expenditure Patterns of the Elderly

Louise A. Heslop, Carleton University
ABSTRACT - a regression-based cohort analysis procedure is used to study the trends in expenditures of the elderly and near elderly between 1969 and 1982. The data derive from the six Family Expenditure surveys conducted during that period. The results suggest that age is less important in predicting expenditures than is usually suggested. Instead the period of birth and the period in which the expenditures are made are far more important. In particular, cohort membership is valuable in understanding trends in many key areas, such as food, clothing shelter and transportation.
[ to cite ]:
Louise A. Heslop (1987) ,"Cohort Analysis of the Expenditure Patterns of the Elderly", in NA - Advances in Consumer Research Volume 14, eds. Melanie Wallendorf and Paul Anderson, Provo, UT : Association for Consumer Research, Pages: 553-557.

Advances in Consumer Research Volume 14, 1987      Pages 553-557


Louise A. Heslop, Carleton University

[The author would like to acknowledge the assistance of the staff of the Social and Economic Studies Division and the Consumer Income and Expenditure Division of Statistics Canada. The analysis presented in this paper is the responsibility of the author and does not necessarily represent any view or policies of Statistic Canada.]


a regression-based cohort analysis procedure is used to study the trends in expenditures of the elderly and near elderly between 1969 and 1982. The data derive from the six Family Expenditure surveys conducted during that period. The results suggest that age is less important in predicting expenditures than is usually suggested. Instead the period of birth and the period in which the expenditures are made are far more important. In particular, cohort membership is valuable in understanding trends in many key areas, such as food, clothing shelter and transportation.


Cohorts are groups of individuals who experience "the same event within the same time period" (Ryder, 1965:845). The term 'cohorts' is most commonly used to describe those who were born at the same time. They experience the same historical events at the same age. The purpose of this paper is to present the results of an exploratory analytical cohort analysis of expenditures data generated by the Consumer Expenditure Survey conducted by Statistics Canada. It focuses on the elderly and near elderly age groups because of the increased interest in recent years in the economic well-being of the older group. Understanding factors affecting their status is made more complex by the commonly experienced co-incidence of aging, change in life position (worker to retiree), and change in income level and source. Moreover, economic policies directed at the elderly population have been changing dramatically over the last decade so those becoming elderly at different times in the recent past have been experiencing substantially different economic environments. Analytical procedures designed to sort out the differential effects of these many interrelated factors seem worthy of examination, application and discussion.

An Overview of the Cohort Analysis Procedure

Different birth cohort groups experience different social, economic, political and technological environments. "Each cohort experiences a different set of environmental events at each stage of its life course than did each preceding cohort when it occupied the same life stage" (Rentz, Reynolds and Stoutz, 1983:12). Cohort analysis attempts to sort out the outcomes of commonalities of experiences in terms of what were the results of the shared experiences themselves and what were the results of the age at which they were experienced. To do this, different cohort groups born at different times but experiencing the same historical events (but at different ages) must be compared. It is

difficult, however, to sort out the complex interactional effects of cohort membership, age and environmental influences. Single cross-sectional studies across different age groups at one point in time cannot unravel these interactions because the effects of age and cohort membership are confounded. So questions regarding behavior patterns which differ across survey respondents in such studies cannot be answered as to whether these differences are attributable to how old the respondents were or to the period in which they were born. Longitudinal studies which use repeated measures of either the same respondents or of sub-samples of the same population (cohorts), do not allow for distinguishing between age effects and the effect of the historical period all respondents shared. Even time-lagged designs comparing samples of the same age from different cohort groups at different times confound the effects of cohort and historical period.

Cohort analysis techniques address the problem of the confounding of variables involved in the efforts to sort out the effects of age vs. cohort vs. period. Essentially with normal procedures it is impossible because of the inherent interrelationship of age, cohort membership and time. For example, at any given point in time, there is only one cohort group at any specific age. To overcome this problem, a series of models was developed in this study for each expenditure category analyzed. The model outcomes were compared for interpretation.

To sort out the effects of each of the three factors under study - age, cohort membership and period - a series of constrained multiple regression models is used. (Mason 1973; Rentz and Reynolds 1980; Rentz, Reynolds and stoutz 1983) They take the form:

Yijk = a + bi + gj + dk + eijk

where Yijk - dependent variable (expenditure level on some category of goods or services)

bi = the effect of the ith period

gj = the effect of the jth age

dk = the effect of the kth cohort

eijk = error term

i, the period, ranges from 1 = 1969 to 4 = 1982

j, the age group, ranges from 1 = 57-60 years of age to 6 = 77-80 years of age

k, the cohort, ranges from 1 = born 1889-1892 to 9 = born 1922-1925

Since all the independent variables are dummy variable conditions, the regression actually reduces to an analysis of variance problem. However, the presentation of the results will use standard regression format for ease of interpretation and explanation.

To permit inclusion of all three sets of factors in the model, the linear dependency is reduced by constraining two effects to be equal at the same time within two of the three variable sets (age, cohort or period). The model results will vary with the choice of the factors constrained and the "suitability of the assumptions can be judged by how well each model fits the data", i.e., by the value of the R2 (Rentz, Reynolds, Stoutz, 1983:14). Thus, for each expenditure category, three models were fitted and identified according to the unconstrained factor:

(1) a "period model" in which the oldest two age groups were constrained to be equal and the earliest two cohorts were so constrained,

(2) an "age model" in which the first two periods were constrained and the earliest two cohorts,

(3) a "cohort model" in which the first two periods and the oldest two ace groups were constrained to be equal.

The choice of pairs of factors to constrain was determined on the basis of data quality and interest. The most recent periods were of the most interest so they were not constrained. Also the earlier periods had substantially lower rates of inflation than the later periods. The earliest cohorts and the oldest age groups were the smallest sample groups, so were combined to improve the stability which affects the reliability of the results. Also, these earliest cohort groups will be the least affected by policy changes in the future as their numbers will decline very rapidly. Combining those over 73 years of age allows for the comparison of those under 65 (non-elderly) to those 65-72 (young-old) to those 73 and over (old-old).

The expenditure categories chosen for analysis included all major categories of expenditures reported in the Family Expenditure Survey publications as well as subcategories of particular interest given the group under study. (For a full explanation of the procedures of these surveys and definition of these categories, see Statistics Canada (1978).)

Analysis of expenditure data in this way can help sort out the sometime cumulative, sometimes counterbalancing effects of age, period and cohort. It can help answer such questions as "Do changes in health expenditures arise because as people age they need more health care services or because health care costs are rising rapidly over time or because people born more recently perhaps demand a higher level of care"? "Do food expenditures decline because as people age they eat less or because people born in earlier times know how to economize better on food purchases"? "Do people spend less on cars as they age because they can no longer drive or because the cohorts born in earlier periods are less used to cars"? "In general, do the elderly spend less because they are aging and perhaps need less because some costs, such as for government subsidized health care, for them are declining over time or because the cohort has less accumulated wealth and pension flows from which to draw"?

The exact reasons for the findings of period, age, or cohort effects, of course, cannot be determined directly from the analysis. Sometimes, the coincidence of historical events and period or cohort effects gives a strong indication of a causal or contributory link. Often theories regarding patterns of aging can contribute to a fuller understanding of the empirical results regarding aging effects.

As with many statistical analysis procedures, this method of cohort analysis is not without controversy. Glenn (1977) is concerned that the linear additive form of the regression model used ignores possible interaction effects. However, the procedure adopted here does allow for at least first level quantitative analysis while other procedures involving only visual inspection and interpretation yield rather vague outputs. More complex interactive term models could perhaps be explored in future work to examine for such interactive effects. Glenn does stress the need for interpretation of results in the light of "outside" information to confirm beliefs about what is observed in the data. This approach has been adopted here. Also the exact numerical level of the parameter estimates have not been stressed, but rather only the general direction of changes noted. The ease of explaining the significant findings provides support for the procedure and its usefulness in verifying trends which might be expected and in suggesting some new areas for explorations. Rodgers (1982a) and Jagodzinski (1984) suggest that the problems of analysis of age, period and cohort effects would be best served by eliminating one of the three variables. Rodgers (1982) feels that the underlying explanatory factors that are associated with the eliminated variable could then be substituted. For example, if the period effect is believed to be the result of period-related economic conditions, then some measures of the economic conditions themselves should be substituted. This argument is particularly appealing if the goal is one of understanding the phenomenon under study, not just mathematically predicting it. However, the initial use of the Mason (1973) procedure can help to identify whether such underlying factors are best sought within age, period or cohort - related phenomenon.


The data used for the regression models were average expenditure levels calculated on a subsample of households interviewed for the 1969, 1974, 1978 and 1982 Family Expenditure SurVeys. The subsample contained households of single unattached individuals and married couples within the age range 57-80 years of age living in the eight cities of St. John's, Halifax, Montreal, Ottawa, Toronto, Winnipeg, Edmonton and Vancouver. The sample sizes varied between survey years but typically were well above 500.

Some changes have been made in the original data categorizations in order to ensure uniformity of classification of expenditures from survey year to survey year. Wherever possible all categories were brought into agreement with 1982 classification procedures.


Table 1 lists the expenditure categories and overviews the analysis findings for both the percentage of total expenditures devoted to each category of expenditures and dollar values of these expenditures. For each expenditure category the best fit model with its significant coefficients, (excluding the intercept term) and their sign are listed. Since all models have the same number of variables, the best fit model is the one with the highest R2. This model in each case is identified with an asterisk in the table.



Percentages of Total Expenditures

In general, the regression models were very good at explaining variations in percentages of expenditures devoted to categories except in the cases of 'Owned Living Quarters', and 'Recreation' expenditures. In these categories the percent of expenditures levels were so variable that no good models were found and no further detailed discussion of these categories of expenditures is included.

With the exception of men's clothing, period and cohort models and not age models are the most explanatory models based on the measure of R2. Cohort models are best in explaining 'Food' expenditures, total 'Shelter' and the shelter subcomponent of 'Rent', 'Transportation', and the transportation subcomponent of 'Car and Truck Purchases', and 'Security' expenditures. Period models excelled in the areas of 'In-home Energy', 'Clothing' and 'Women's Clothing' and 'Medical and Health', 'Gifts and Contributions' and 'Taxes' expenditures.


The data for this category has been plotted to aid in explaining the results and how they can be interpreted. Two graphs have been made of the data, one with lines joining the same age groups across years to highlight age factors (Figure 1) and the other doing the same for the cohort groups (Figure 2). Referring to Table 2, it can be seen that the cohort models were best. However, within these models, individual cohort variables did not emerge as important. In Table 1 it can be seen that the cohort groups do appear to have some pattern in the series of levels ranging from cohort 1 at the top with one data point, down to cohorts 8 and 9 at the bottom. However, there is quite a bit of overlap preventing any significant individual differences. Despite this overlap, there is a declining trend in percentage of total expenditures spent on food with later born cohorts. The significant differences from the base conditions were found mainly among the age ranges. In the cohort model the two oldest groups are constrained to be equal. As can be seen in Figure 1, this is an appropriate choice. The age differences show up dramatically, but not as individual age groups. Rather, there are three distinct groups - those over 73, the old-old; those 65-72, the young-old; and those under 65. However, the biggest jump is between those under 65 and those over. Age groups 1, 2 and 4 were found to have significantly lower percentage expenditures on food than the oldest two groups. Age 3 (65-68) probably failed to meet the criteria because of its somewhat greater variance and the very close proximity of its 1974 level to that of the oldest groups. A common sense interpretation of the plots suggests that there are real differences in a consistent way between the three age levels with the major shift occurring after age 65. Consumption changes in this category come immediately upon retirement. There is a dramatic shift upwards in budget allocations to food. Those under 65 years of age spent four to eight percent less on food than seniors. Food is perhaps the key indicator of well-being. This sudden shift in percentage of expenditures devoted to this category clearly indicates a drop in living standard.





In terms of period effects, the last period shows a decline for almost all age and relevant cohort groups. Lower increases relative to incomes in food prices in the early 1980's would account for this drop. All groups spend significantly less as a percent in 1982 than in the base period, suggesting an improvement in living standards probably resulting from lower increases in food prices in the early 1980's relative to income.

No cohort groups were significant. There did not appear to be any impact on this category due to the period in which people had been born, raised, and raised their families (for example, no Depression mentality). However, the significance of the overall model and the trends noted in Figure 2 suggest that there is a trending down for cohorts in the proportions of budgets they have to spend on food. Why might this be? It could not be simple an effect of absolute age-linked individual income levels or price levels. These would show up as age or period effects. The explanation has to lie in something with sustained impact on the whole cohort. This may be a lifetime standard of living or relative income which has been rising with each cohort. Food does appear to be a key indicator of the impact of such shifts.


The overall 'Shelter' expense includes the categories of 'Rented Living Quarters' expenditures, 'Owned Living Quarters' expenses (mortgage interest, taxes, repairs and maintenance>, and also expenditures for lodging away from home and utilities expenses. Overall 'Shelter' expenditures are best explained by the cohort model. All cohort groups spent significantly less than the base cohort group. An examination of the model coefficients (not reported here) it is almost an uninterrupted negative trend (see Heslop 1985). Also, the two periods in this model are significant, showing increasing percentages of total expenditures devoted to all shelter areas in the 1978 and 1982 period than in earlier periods. The very high interest rates during this period may be the cause of this shift, was well as rapid rises in utility costs.

No models were significantly useful in explaining variances in the 'Owned Living Quarters' expenditures. Rented living quarters expenditures, however, are well explained by the cohort model. Like the total 'Shelter' model, the cohort model notes a significant and generally monotonic decline right from the 2nd to 9th cohort groups. These findings combined with the overall shelter negative cohort effects and the positive (though non-significant) cohort coefficients for 'Owned Living Quarters' suggest that successive cohort groups of the elderly have been gaining in their ownership of housing stock over the years, another indication along with food expenditures of cohort improvements.

'In-Home Energy' seems to be better explained by the period model but no significant factors emerge. The tendency is to see an increase in all periods but none reach significance. The rapid rises in energy costs are probably responsible for the effect. Matching increases in incomes likely prevent any significant increases in percentage expenditures from showing up.


'Total Clothing' and 'Women's Clothing' are best explained by period models. Both contain significant negative period coefficients for 1974 and 1982 and positive cohort coefficients for those born around the turn-of the century. The general cohort trend is upward but only the first two are significantly different from the base period. Although no coefficients are significant, 'Men's Clothing' was best explained by the age model. So while there is a tendency for women to be altering their clothing patterns with each succeeding cohort, men still seem to be locked into a pattern of a decline with age in the proportion of budgets spent on clothing.

Medical and Health Care

'Medical and Health Care' expenditures include contributions to government- sponsored and private health insurance programs as well as direct fees to medical professionals, hospitals and the purchases of medicines and drugs, and supplies and services. The most successful model by far is the period model. In holding only the 1969 period as the base period, it allows for the capturing of the effects of the growth of coverage of government health insurance programs in the early 1970's. This can be seen in comparing the three models. The age and cohort models equate effects of the 1969 and 1974 periods and note positive but non-significant coefficients in subsequent time periods. However, the period model uses only 1969 as the base year period. A dramatic decline in percent expenditures in all periods is noted. The introduction of government subsidized health care has had a dramatic effect for all age groups, significantly decreasing budget allocations.

It is worth at least observing that the difference between the 1969 and the 1978 and 1980 periods is less significant than the difference between 1969 and 1974. The higher variance associated with the latter two survey years seems to be the limiting factor, since the absolute values of the coefficients increase over time. The introduction of user fees and extra billing by some doctors may be the cause of this greater variance in the more recent years. Finally, although there is a negative trend in budget allocations with decreasing age, no age effects were significant.


The cohort models outperform the other two in the transportation categories. In both the overall 'Transportation' expenditures category and especially in the 'Automobile and Truck Purchase' component, all the significant variables are cohort ones. For the purchase variable, all cohorts born in this century spent a higher percentage on car and truck expenses than did the earliest cohort group. The car is a twentieth century phenomenon. The children of this century developed it and are using it. The cohort pattern is monotonic and strong with only a slight drop in the latest born cohort' group, perhaps suggesting some levelling off.

Taxes and Securities

Analyses of tax expenditures suggest that the period model explains the data best. There is a substantial jump in 1982. Also three age variables are significant. Generally, those under 68 years old spent more than the oldest two groups in the period model. It would appear that taxes do not drop immediately upon reaching age 65, but rather that there is a short lag. The cohort model excels in explaining total 'Security' expenditure percentages but no variables are significant within this model


'Gifts' expenditures as a percentage are best explained by a period model in which some coefficients in all categories are significant. Successively larger percentages of total expenditures have been given away with each period after 1969. The youngest age group gives significantly more than the oldest group, and there is a general decline which can be seen in the coefficient values with increasing age. Finally there is a very strong tendency for later born cohort groups to give less and less than earlier born ones. Here the value of cohort analysis can be most readily seen in separating out the opposing forces of period from cohort and age factors. As people age and as later born cohorts mature, a smaller percentage of budgets is given away. Upward increases are related to the period of time through which we have come. If there is a change in the period-related factor which is responsible for this period effect (perhaps it is rising incomes) then declines in contributions can be expected. Declines in percentages given away may reflect shifts over time in who is perceived to be responsible for charitable activities and the support of family members.

Dollar Expenditures

Period effects models are much more likely to dominate in explaining variances in dollar expenditures. The mid to late 1970's was a period of rapidly rising prices. Incomes also rose rapidly throughout much of the 1970's. So, large differences in expenditures are seen from one survey year to the next because both prices and incomes were rising rapidly. The effect is especially strong in the period models where the comparison year is the one single pre-1970's year, as opposed to other models where 1974 and 1978 are compared to 1969 and 1972 combined.

As a result a quick summary can be given by saying all expenditures categories had at least one significant model. This was the period model in all cases but 'Clothing' categories, 'Car and Truck Purchases', 'Recreation', 'Taxes', and 'Security' expenditures which were best explained by cohort models. No age models emerged as superior.

Period effects were significant in all best models except for 'Total Clothing' and 'Men's Clothing', 'Transportation', 'Car and Truck Purchases' and 'Recreation' expenditures. Cohort effects also were very important as explanatory variables, frequently entering as long strings suggesting a substantial birthdate-related change in consumption patterns. This can be seen for 'Clothing' and especially 'Men's Clothing', 'Car and Truck Purchases', 'Recreation' and 'Personal Taxes'. Age effects were only significant in 'Recreation' expenditure. All the significant factors are positive indicating increases over time, higher expenditures by each successive cohort, and, where age does appear, higher expenditures by younger groups.

Examination of the plots of the data (not reported here due to space constraints, see Heslop 1985) reveal important information. The very strong period effect is obvious in both the age and the cohort plots. It is strong and consistently felt in every period as is indicated in Table 1- The only other statistically significant factor is that for cohort 8. Only this one later born cohort is identified as spending significantly more than the earlier cohorts. One might wonder if this is a fluke. However, examination of the groups suggests the identifying of the once cohort as anomalous and the only one to shift upwards in expenditures for food may be an overly conservative conclusion. A very consistent and almost monotonic up,ward cohort shift is noticeable from the almost parallel set of lines. Each succeeding cohort is spending more in dollar terms than earlier ones.

The same evidence suggests that age trends are there although not statistically significant. The year-to-year patterns across all groups are generally similar. Therefore, the same factors (price rises, income increases) are likely affecting all groups about the same. Age is not the major determinant of food expenditures over time and across succeeding cohort groups. In other words, age is not a good predictor of need.


In final assessment of the relative effects of age cohort - period on expenditures four observations are made:

1. Percentage expenditures are useful, perhaps even more useful to look at than dollar expenditures, unless inflation is relatively low. Non-inflation period effects and cohort and age effects were overwhelmed in the 1970's by the dominance of double-digit inflation. and the resulting large period effects in dollar expenditure cohort analysis. Percentage expenditures analysis can provide a relative and perhaps more relevant analysis in this case. Attempts to control for and factor out inflation effects would really necessitate the existence of a set of accurate, credible, age-related price inflation measures, which are not currently available.

Percentage figures give a picture of the division of the resources available. They remove again the very powerful effect of a great disparity of incomes and total expenditures among the youngest and oldest groups at any one time and among different cohorts at any one time. At the same time they capture the effect of this budget size in how the different groups divide up their different sized pies.

2. Period effects are important and require careful treatment. If key changes in the marketplace occur, the set-up of the cohort analysis models should be sensitive to these changes. The selection of the base periods should avoid combining two very unlike periods, which may result in the masking of the significant break point.

3. Age is not as powerful a factor as might originally be expected. It appears useful in any consistent way only for taxes as a percentage of expenditures. Age alone explains very little of how, on the aggregate, people allocate their monies. These findings should provide a clear signal of the need to reassess commonly held assumptions that the old need less. Moreover, where age factors are important within models, e.g., regarding food as a percentage of expenditures, there are clear indications of changes in well-being upon retirement.

4. cohort models play a Significant part in explaining overall variance in some key macro-expenditure areas of food and shelter in percentage terms. In dollar terms they emerge for clothing, car and truck purchases and recreation which may suggest that yesterday's image of the elderly may be rapidly becoming outdated with the emergence of a new elderly who are more interested in leading full, active lives and have the monetary resources to do so. This has implications for future income needs of the elderly. It is no longer acceptable to say the elderly spend less because they do less and wish to do less. The income needs of the new elderly are expanding.


Baltes, Paul B. (1968), "Longitudinal and Cross-Sectional Sequences in the Study of Age and Generation Effects", Human Development, 11, 145-71.

Blalock, Hubert M. (1966, "The Identification Problem and Theory Building: The Case of Status Inconsistency", American Sociological Review, 31, 62-61.

Glenn, Norval D. (1977), Cohort Analysis, Sage University Paper Series on Quantitative Applications in the Social Sciences, 07-001, Beverly Hills and London: Sage Publications.

Havens, Betty (1980), "Differentiation of Unmet Needs Using Analysis by Age/Sex Cohorts" in Aging in Canada! Social Perspectives, Victor W. Marshall, ed., Don Mills, Ontario: Fitzhenry and Whiteside.

Heslop, Louise (1985), "Cohorts - Going through Life Together: An Analysis of the Expenditure Patterns of the Elderly", Research Paper No. 13A, Social and Economic Studies Division, Statistics Canada, Ottawa.

Jagodzinski, Wolfgang (1984), "Identification of Parameters in Cohort Models", Sociological methods and Research.12,4,375-398.

Mason, Karen O., William M. Mason, H.H. Winsborough, and W. Kenneth Poole (1973), "Some Methodological Issues in Cohort Analysis of Archival Data", American Sociological Review, 83, 242-58.

Rentz, Joseph O. and Fred D. Reynolds (1980), "Separating Age, Cohort and Period Effects in Consumer Behavior", in Advances in Consumer Research, Vol. 8, Kent B. Monroe, ed., Washington, D.C.: 596-601.

Rentz, Joseph O., Fred D. Reynolds and Roy G. Stoutz (1983), "Analyzing Changing Consumption Patterns with Cohort Analysis", Journal of Marketing Research, XX, 12-20.

Riley, Matilda White (1973), "Aging and Cohort Succession: Interpretations and Misinterpretations", Public Opinion Quarterly, 37, 35-49.

Rodgers, Willard L. (1982a),"Estimable Functions of Age, Period, and Cohort Effects", American Sociological Review, 47,774-787.

Rodgers, Willard L. (1982b), "Reply to Comment by Smith, Mason, and Feinberg", American Sociological Review, 47,793-796.

Schaie, K. Warner (1965), "A General Model for the Study of Developmental Problems", Psychological Bulletin, 64, 92-107.

Smith, Herbert L., William M. Mason and Stephen E. Feinberg(1982)"More Chimeras of the Age-Period-Cohort Accounting Framework: Comment on Rodgers" American Sociological Review, 47,787-793.

Statistics Canada (1978), "Urban Family Expenditures, 1978" Vol. 3. Catalogue Number 62-551.