2.1 Conceptual framework

Empirical evidence of the retirement-consumption puzzle raises doubt to the standard LCH of consumption stating that households save their income while working in order to smooth consumption in response to an anticipated income decline at retirement. The extended version of the LCH including the effects of liquidity constraints and uncertainty helps explain the substantial consumption drop around retirement. However, the analysis still focuses on the consumption of market goods since the LCH assumes that consumption and leisure are separable in utility, which implies a constant marginal utility of wealth. If preferences are non-separable and consumption and leisure are substitutes, the marginal utility of wealth will fall as leisure time increases with retirement. This potentially results in the significant decline in consumption expenditures [Banks et al. (Reference Banks, Blundell and Tanner1998), Hurd and Rohwedder (Reference Hurd and Rohwedder2013)].

To empirically investigate the effect of retirement on changes in household consumption patterns in Thailand, this paper follows the life-cycle model augmented with home production. The formal model of consumption and time allocation was initially developed by Becker (Reference Becker1965). A representative household acts as both a consumer and a producer and is assumed to combine market expenditures and time to produce consumption commodities to maximize utility subject to relevant constraints. The key implication of the model is that the decrease in the opportunity cost of time upon retirement may induce retired households to engage more in-home production previously purchased. Therefore, the inclusion of consumption, production and time allocation enables one to better examine consumer behavior of the elderly in which they will become the largest population group in the near future.

2.2 Empirical analysis

In an attempt to identify the causal impact of retirement on household consumption and income around retirement, the following specifications, which is extensively used in the literature [e.g., Cho (Reference Cho2012), Hori and Murata (Reference Hori and Murata2019), Been et al. (Reference Been, Rohwedder and Hurd2021)], are estimated:

(1)$$\ln ( {C_{it}} ) = \alpha _c + \beta _cR_{it} + {\boldsymbol Z}_{{\boldsymbol it}}{\boldsymbol \gamma }_c + {\rm \epsilon }_c$$

(2)$$\ln ( {Y_{it}} ) = \alpha _y + \beta _yR_{it} + {\boldsymbol Z}_{{\boldsymbol it}}{\boldsymbol \gamma }_y + {\rm \epsilon }_y$$

where C _it denotes the different measures of consumption expenditures of household i in year t, and Y _it stands for household i's income in year t. The key explanatory variable R _it is a dummy variable for the retirement of the male household head, equal to one if retired and zero otherwise.Footnote ⁴ The vector Z_it controls for a head of household and household characteristics, including household size, the household head's years of education, gender and age, and dummy variables for regions and urban and rural areas, and the error terms ε_c and ε_y are assumed to be normally distributed with zero mean and finite variance.

The coefficients β _c and β _y measure the impacts of retirement on household consumption and income, respectively. One would expect a negative sign of β _y coefficient resulting from a sharp decline in household income as the head retires from paid work. Furthermore, the LCH predicts that β _c should not be significantly different from zero. In contrast, if the β _c coefficient is substantial and negative, it provides supportive evidence of the retirement-consumption puzzle. As described in Hurst (Reference Hurst2008), the sign and magnitude of β _c may vary across the consumption outcomes, and the large decline in consumption expenditures upon retirement are generally concentrated on work-related and food categories. To explore these possibilities, five expenditure categories are separately examined: (1) total expenditure; (2) food-at-home; (3) work-related expenses (on transportation and communication, clothing, and food consumed away from home); (4) non-durable entertainment; and (5) other remaining expenses. An additional set of regression analyses is also conducted for the subsample to investigate differences in expenditure changes across socioeconomic groups.

To assess whether and the extent to which retirement leads to changes in time devoted to home production among retirees, the resulting regression model is

(3)$$T_{it} = \alpha _t + \beta _tR_{it} + {\boldsymbol Z}_{{\boldsymbol it}}{\boldsymbol \gamma }_t + {\rm \epsilon }_t$$

where T _it is the individual's amount of time spent on a specific home production activity such as food preparation, house cleaning, and shopping, and ε_t is the error term. One would expect to obtain a positive sign of β _t, which reflects the increased hours spent on home production after retirement due to a decrease in the opportunity cost of time. However, it should be noted that the impact of retirement on time use patterns may differ across activity types and genders.

One important estimation issue is that the ordinary least squares (OLS) estimates of Equations (1) to (3) may suffer from endogeneity bias because the retirement status is not exogenous, which is often correlated with the error terms. To obtain consistent estimates of the impact of retirement on consumption and time use patterns, an instrumental variable (IV) approach is applied using the Thai legislation that the mandatory retirement age for public sectors and the minimum retirement age for private sectors with eligibility for the severance pay are set at the age of 60. Following recent literature, this paper uses a dummy variable indicating whether the household head's age is 60 or over as an instrumental variable (IV) for the retirement status of the male household head [Battistin et al. (Reference Battistin, Brugiavini, Rettore and Weber2009), Moreau and Stancanelli (Reference Moreau and Stancanelli2015), Li et al. (Reference Li, Shi and Wu2016), Been et al. (Reference Been, Rohwedder and Hurd2021)]. In this specification, it does not require that all individuals are retired at age 60; indeed, they may retire before or after age 60. However, it is expected that the probability of being retired increases significantly at the threshold age of 60 with a value between 0 and 1.

To address the potential endogeneity problem, the two-stage least squares (2SLS) method is employed to estimate the casual impact of retirement on household consumption and home production. The first-stage equation using a probit model is given by

(4)$$R_{it} = \delta _0 + \delta _1D_{it} + \delta _2f( {s_{it}-60} ) + {\boldsymbol Z}_{{\boldsymbol it}}{\boldsymbol \delta }_3 + {\rm \epsilon }_r, \;\;R = 1[ {s_{it} \ge 60} ] $$

and the second-stage equation is

(5)$$\ln ( {C_{it}} ) = \alpha _c + \beta _c\hat{R}_{it} + {\boldsymbol Z}_{{\boldsymbol it}}{\boldsymbol \gamma }_c + {\rm \epsilon }_c$$

where s _it is the male head's age of household i at year t, f(s _it − 60) is a polynomial function of order p in age, f(s _it − 60) =  (s _it − 60) + (s _it − 60)² + … (s _it − 60)^p to make the functional form more flexible, D _it is the dummy variable indicating the household head's age is 60 or over in year t, and $\hat{R}_{it}$ is the predicted probability of the household's head retirement.Footnote ⁵ Subsequently, the above estimation procedure is repeated to examine the effect of retirement on home production in that the dependent variable of Equation (5) is replaced by the amount of time allocated to home production (T _it).

Turning to the parameters of interest, the expected percentage change in spending for a particular category at retirement can be computed from $e^{\beta _c}-1$, and the additional time spent on home production is β _t. More precisely, the IV estimates of β _c and β _t represent the local average treatment effects (LATE) of retirement on changes in consumption expenditure and time spent on home production for the compliers, defined as individuals who are already retired if their age reaches 60 and those are not retired if their age is less than 60. This estimation strategy is equivalent to fuzzy regression discontinuity design (FRD) in which the FRD estimator can be specified as:

(6)$$\beta _o^{FRD} = \displaystyle{{\mathop {\lim }\limits_{\varepsilon \to 0} E[ {O_i\vert s_i = 60 + \varepsilon } ] -\mathop {\lim }\limits_{\varepsilon \to 0} E[ {O_i\vert s_i = 60-\varepsilon } ] \;} \over {\mathop {\lim }\limits_{\varepsilon \to 0} E[ {R_i\vert s_i = 60 + \varepsilon } ] -\mathop {\lim }\limits_{\varepsilon \to 0} E[ {R_i\vert s_i = 60-\varepsilon } ] }}$$

where $\mathop {\lim }\limits_{\varepsilon \to 0} E[ {R_i\vert s_i = 60 + \varepsilon } ] -\mathop {\lim }\limits_{\varepsilon \to 0} E[ {R_i\vert s_i = 60-\varepsilon } ] \ne 0$ and O _i is the outcomes of interest including consumption expenditure, household income, and time spent on home production.Footnote ⁶

However, it is worth mentioning the key assumption for the validity of the regression discontinuity model that the mean value of consumption conditional on the household head's age is continuous at the age threshold of 60. In other words. no other confounding factors can trigger a discontinuity in household consumption, except for one's retirement status. To test for the validity of this assumption, regression analysis is implemented where the dependent variable is each predetermined characteristic, including household size and the household head's years of schooling, and the explanatory variables consist of the retirement status and controlled household characteristics [see Battistin et al. (Reference Battistin, Brugiavini, Rettore and Weber2009), Li et al. (Reference Li, Shi and Wu2016)]. The insignificance of the retirement coefficient indicates that the regression discontinuity design (RDD) continuity assumption is satisfied. Further, the McCrary (Reference McCrary2008) manipulation test for the running variable is performed to ensure the robustness of the testing results, as done in Stancanelli (Reference Stancanelli2017). The idea is that the density of the household head's age should be continuous around age 60 if there is no manipulation, providing evidence in favor of the RDD continuity assumption. Another concern is that the running variable used in this paper – household head's age in years – is discrete, so the appropriate polynomial function of s _it is added to the retirement equation in order to obtain consistent estimates of the LATE. See Dong (Reference Dong2015) and Dong and Yang (Reference Dong and Yang2017) for further details. Finally, heteroscedasticity-robust standard errors of all estimates will be used for hypothesis testing throughout the paper.

3.1 Data

Two data sets are used in this paper: Socio-Economic Survey (SES) and TUS. The two surveys are nationally representative and cover all provinces in Thailand, both conducted by National Statistical Office (NSO). The analysis of changes in household consumption and income around retirement relies upon the SES data gathered in 2013, 2015, 2017, and 2019. The SES is a large cross-sectional survey, which provides detailed information on income, expenditures, household composition, and geographical and socioeconomic characteristics. In addition, the SES provides information on household assets and debt, debt payment, migration, and private transfers. The survey data were also collected at the individual level, including the level of education, occupation, employment status, wage income, age, and gender. A stratified two-stage sampling procedure was implemented for the SES: the primary sampling unit was blocked for municipal areas and villages for non-municipal areas, then private households were selected within these blocks and villages. There were roughly 40,000 sampled households in each survey.

For the sake of investigating household expenditure patterns around retirement, this paper focuses on a subset of male-headed households aged 50–70 in each survey year. The final sample includes 51,518 households for four cross-sections of the SES (n = 11,926 in 2013; 12,659 in 2015; 13,063 in 2017; and 13,870 in 2019). It should be noted that the SES collects age in years, not the year of birth; thus, it is potentially affected by recall error since the age is likely to round to the nearest integer at the time of the survey.Footnote ⁷ Based on individuals' current working status question, retirement status is defined as the male head reported himself at the interview date as being retired or economically inactive for a number of reasons like illness, disability, not willing to work, or elderly.

Household expenditures consist of spending on food, non-food items, durables, and services, with comprehensive details for each category. A face-to-face interview was conducted with the household head or member of the household at the respondent's premises using retrospective questions to gather information on household consumption and expenditure with three different recall periods: the previous week (e.g., food items and beverages), the previous month (e.g., residential rent, clothing, and transportation), and during the past 12 months (e.g., educational expenses, vehicle repair and maintenance costs, which are purchased less often during the year). The average response rate was 74.8% at the national level. In this paper, five household expenditure categories are analyzed: total expenditure, food consumed at-home, work-related costs, non-durable entertainment, and others.Footnote ⁸ Expenditures on food-at-home include food purchased and received-in kind (but excluding alcoholic beverages and tobacco products). Expenditures on work-related items include food consumed away from home, clothing and personal services, and transportation and communication (not including vehicle purchases). Non-durable entertainment expenditures comprise of vacations, outdoor activities, games, sport equipment, and other recreation and religious activities. Other remaining refer to expenditures on items not already accounted for in the above. All reported income and expenditure data are calculated on the monthly basis and adjusted to constant 2019 prices using the consumer price index (CPI) from Thailand's Ministry of Commerce.

Echoing recent studies, not only do retirees have more leisure time to engage in home production like cooking or cleaning, but they can also purchase most goods and services less costly by spending more their leisure time on searching and shopping to find good bargains. Therefore, the observed decline in household expenditures after retirement might not necessarily imply lower living standards. This paper also uses the 2015 TUS of Thailand to ascertain whether the casual relationship between individuals' retirement status and time spent on home production activities. Stratified three-stage sampling was used for the TUS in that Bangkok and other provinces are strata. Each stratum except for Bangkok is divided into two administrative areas: municipal (urban) and non-municipal (rural). The primary sampling unit was an enumeration area (EA); the secondary sampling unit was private households within the EA; and the tertiary sampling unit was an individual aged 6 years and over from each sampled household, which was selected through simple random sampling. The initial sample consisted of 83,880 individuals from different households.

The analysis here restricts to the sampled individuals aged 50–70 (also age in years), thus yielding a sample of 23,432 individuals (28%): 10,402 males and 13,030 females. Respondents' retirement status is defined as similar to that in the SES since the respondents from the two surveys were asked the same question on their employment status. From July 2014 to June 2015, personal interview and self-completed questionnaire were conducted to collect information on how the respondents spent their time doing different activities in every 10 min of a 24-h day. The overall response rate was exceptionally high for the TUS survey (93.8%). However, a further point worth noting is that the TUS provides rather limited information on household demographical and socioeconomic characteristics, as compared to the SES data.

The classification of activities for the TUS of Thailand is primarily based on the United Nations Statistics Division (UNSD). There are eight major divisions of activities: (1) paid work, (2) unpaid domestic services, (3) unpaid caregiving services to household members, (4) community services, (5) learning, socializing, (6) community participation and religious activities, (7) leisure and sports, and (8) personal care and maintenance. The amount of time spent on the above-mentioned activities is measured by minutes per day. Following recent literature examining the retirement effect on home production [e.g., Stancanelli and Van Soeast (Reference Stancanelli and Van Soeast2012), Atalay et al. (Reference Atalay, Barrett and Staneva2020), Been et al. (Reference Been, Rohwedder and Hurd2021)], activities are grouped into the following categories:

1. (1) Food management (preparing and serving meals, cleaning up afterwards, and other related activities)

2. (2) Cleaning of dwelling and surroundings

3. (3) Maintenance and small repairs and care of textiles (washing, drying, ironing)

4. (4) Shopping

3.2 Summary statistics of the sample

Figure 1 illustrates the relationship between the proportions of retired male heads and individuals and their age computed from the SES and TUS data, revealing a significant positive increase in the proportion of retirement at age 60. Moreover, Table 1 presents the structure of household expenditure for the sampled households with their male-heads aged 50–70, calculated from the four SES surveys over the period 2013–2019. The percentage shares of the total consumption of each category are in parentheses. The average total monthly household expenditure has slightly decreased by 1.4% from 21,083 baht in 2013 to 20,779 baht in 2019. Expenditures on work-related items account for almost one-fourth of total spending during such period, of which more than half are transportation and communication expenses. The average share of household expenditure spent on food consumed at home remains unchanged at almost 30% of total spending. Additionally, the share of food consumed away from home consumption exhibits a higher fluctuation compared to other expenditure categories, with an average of 6%. The average household spends around 11–12% of total household spending on entertainment. As can be seen, the average expenditure on three broad categories (work-related, food consumed at home, and entertainment) is 13,937 baht per month, which account for almost two-thirds of total household spending.

Figure 1.

Proportion of retired households and individuals by age (a) Retired households, SES 2013–2019 (b) Retired individuals, 2015 TUS.

Table 1.

Monthly household expenditure structure

Notes: Own computations using data from the SES surveys of Thailand for households with male heads aged 50-70. The consumption expenditures are expressed in constant 2019 Thai baht. The percentage shares of each consumption category are in parentheses.

Based on the SES surveys in 2013, 2015, 2017, and 2019, the average monthly household income and expenditure are 29,219 baht and 21,492 baht, suggesting that the saving rate of the sampled households is quite high, about 26% of household income. The average head's age is about 59 years, of which 44% aged 60 and over, while only 26% reported that they are being retired. A further investigation shows that the proportion of retired heads is quite high (roughly 60%) for households living in municipal (urban) areas. In addition, almost 90% of the sample are married and living with their spouse, 56% lived in municipal areas, 68% completed primary education, and the means years of schooling is 7 years. In terms of demographic characteristics, it finds that the average household size is 3.2, and the number of children (age 15 and under) is 0.5 per family.

Using the 2015 TUS data, the average age of respondents (aged 50–70) is 59 years, which is comparable to that from the SES data. Moreover, 44% of the sample are males, nearly one-third report themselves as being retired, and more than 90% are the heads of the household or their spouses. On average, the respondents spend more than an hour and a half per day on food-related activity. It is likely that retirees replace food consumed away from home with food at-home production. In addition, they spend slightly more than an hour daily on house cleaning (62 min) and maintenance and small repairs (73 min), while the lowest average time spent per day belongs to shopping (41 min). Additional details on summary statistics, obtained from the SES and TUS surveys, are provided in Tables A1 and A2 in the Appendix.