Revisiting the consumption puzzle at retirement

Recent demographic change and rising life expectancy are reshaping population age structures and increasing the share of older individuals. These trends have profound implications not only for pension systems and health expenditures but also key macroeconomic variables such as consumption, saving and economic growth. Yet the direction and magnitude of these effects remain an open empirical question.

The canonical life cycle/permanent income hypothesis (LCPI) predicts that individuals smooth consumption over time by anticipating predictable income changes, including retirement. Under this framework, retirement should not induce a discrete change in consumption, as post-retirement consumption is optimally planned. However, starting with the seminal contribution of Banks et al. (1998), a large empirical literature documents a significant decline in consumption at retirement, a phenomenon commonly referred to as the retirement-consumption puzzle.

This paper explores consumption dynamics around retirement for Spanish households, following the approach of Banks et al. (1998). We test two central hypotheses. First, consistent with life-cycle theory, consumption should evolve smoothly around retirement, without discontinuities. Second, if the LCPI holds, then retirees’ consumption should be insulated from contemporaneous pension reforms. To evaluate these hypotheses, we estimate a consumption Euler equation and assess how consumption growth responds to retirement status and pension policy changes.

Our analysis draws on several Spanish household microdata sources with rich longitudinal dimensions. We use the Encuesta Permanente de Consumo (EPC, 1977–1983) and the Encuesta Continua de Presupuestos Familiares (ECPF, 1985–1996), which follow households for up to 12 and 8 quarters, respectively. As a robustness exercise, we also examine the Encuesta de Presupuestos Familiares (EPF, 2016–2022). These data allow us to exploit within-household variation over relatively long horizons and to address endogeneity concerns using lagged instruments.

The results provide robust evidence of a consumption drop at retirement in Spain, comparable in magnitude to findings for other countries but declining over time. The estimated consumption dip ranges from 10% to 13.5% in 1977–1983, to 4.4%–5% in 1985–1996 and to 1.8%–2% in 2016–2022. Importantly, we disentangle the effect of retirement itself from contemporaneous pension reforms, showing that the two forces often operate in opposite directions and may partially offset each other in empirical estimations.

This paper contributes to the literature in several ways. First, it exploits panel data sets with a longer time dimension than those commonly used in related studies, such as the US Consumer Expenditure Survey (CEX) or the Panel Study of Income Dynamics, enabling more credible identification of retirement effects. Second, the Spanish experience between 1977 and 1996 provides a quasi-natural experiment: retirees experienced rising real incomes, implying that observed consumption declines cannot be mechanically attributed to income losses. Third, by spanning periods of markedly different institutional development, from a regulated economy to a fully modern Eurozone member, we document a gradual attenuation of the retirement consumption drop, consistent with improved financial markets and institutions. Finally, this paper provides structural estimates of the intertemporal elasticity of substitution (IES), with implications for both macroeconomic modelling and pension policy design.

The remainder of this paper is organized as follows: Section 2 reviews the related literature. Section 3 describes the Spanish pension system. Section 4 outlines the theoretical framework and econometric strategy. Section 5 presents the data. Section 6 reports the main results, and Section 7 concludes.

Since the seminal contribution of Banks et al. (1998), a large empirical literature has documented a decline in consumption at retirement, challenging the smooth consumption path predicted by the LCPI^[1]. Proposed explanations can be broadly grouped into two main mechanisms: lifestyle changes associated with retirement and imperfect retirement planning. Given the existence of surveys on this topic, this section provides only a brief overview of these two frameworks ^[2].

One strand of the literature attributes the consumption drop to changes in expenditure needs following labor market exit. Retirement eliminates work-related expenses such as commuting, meals away from home and professional clothing (Aguiar and Hurst, 2013). While these reductions may account for part of the observed decline, their quantitative importance remains debated.

More generally, retirement alters time allocation. With non-separable preferences between consumption and leisure, individuals may substitute market expenditure with time-intensive activities, including home production or more efficient shopping. Empirical evidence supports this channel, showing that increased leisure can partially offset lower monetary expenditure (Luengo-Prado and Sevilla, 2013; Atalay et al., 2020).

A second explanation emphasizes misalignment between expectations formed before retirement and realized post-retirement outcomes. On the income side, individuals may overestimate future pension benefits or their ability to finance consumption from accumulated wealth, especially in systems characterized by institutional complexity or limited financial literacy (Banks et al., 1998; Bernheim et al., 2001; Bucher-Koenen and Lusardi, 2011). Housing wealth, a dominant savings vehicle in many European countries, may further constrain optimal decumulation because of illiquidity (Christelis et al., 2021; Suari-Andreu et al., 2019).

Uncertainty about future expenditures, particularly health-related costs and longevity, also plays a role, as these risks tend to be underestimated prior to retirement (De Nardi et al., 2016; Banks et al., 2019). Behavioral explanations reinforce this view, as present-biased individuals may postpone necessary consumption adjustments until retirement actually occurs (Pagel, 2017).

More recent work highlights substantial heterogeneity in retirement consumption responses, with larger declines among specific income groups (Hurd and Rohwedder, 2013; Marini, 2024). Another emerging strand studies the effects of pension reforms on consumption and savings (Kolsrud et al., 2024), complementing the extensive literature on labor supply responses. However, empirical evidence linking pension reforms to consumption behavior at retirement remains limited.

For Spain, such analyses are largely absent. An exception is Párraga Rodríguez (2022), who finds a high marginal propensity to consume out of pension reforms using a different empirical approach. Nevertheless, a systematic examination jointly addressing the retirement-consumption puzzle and the role of pension reforms is still missing.

This paper fills this gap by studying consumption behavior around retirement in Spain while explicitly accounting for contemporaneous pension reforms. In doing so, we test two core implications of the LCPI: that consumption should remain smooth at retirement and that pension reforms should not affect the consumption of already retired individuals.

Between 1977 and 1996, Spain underwent socio-economic transformations that deregulated the economy, established institutional frameworks and consolidated the pension system. Under the Spanish pension scheme, workers generally qualified for old-age pensions at 65 years (early retirement being possible in some cases,  Appendix 1).

The system featured high replacement rates ^[3] and guaranteed substantial minimum pensions for individuals with low earnings and/or short periods of contribution, which eventually converged with minimum wages. Despite its complexity and group-specific rules, the system ensured generous annual increases; for instance, average monthly pensions rose 3.72-fold between 1982 and 1996 (Ministerio de Inclusión, Seguridad Social y Migraciones, 2023).

Our analysis covers 14 reforms across two subperiods: 1977–1983, characterized by annual adjustments, often below inflation except in 1981 and 1983, and 1985–1996, which included six major reforms ^[4]. To ensure long-term sustainability, a 2011 reform tightened eligibility, gradually raising the statutory retirement age to 67 years by 2027. Consequently, the short-run impact of this recent reform remains limited.

Our theoretical framework is a standard life-cycle consumption model ^[5]. Households are assumed to maximize [1]:

where t is time, T lifetime of the household, E_t expectation operator conditional on the available information at time t, $β∈(0,1)$ discount factor, U utility, c_t consumption over the period indexed by t and $θt$ a vector of demographic and socioeconomic variables (age, education level, family size and labor market status) shifting utility.

Following Banks et al. (1998) and Alvarez‐Cuadrado et al. (2016), we assume the following functional form for $θt:$

The additively separable utility function is Constant Relative Risk Aversion:

where σ denotes the coefficient of relative risk aversion. As our primary interest lies in the behavior of consumption, we abstract from a detailed characterization of production. Instead, we assume a simple endowment economy where households save through a single asset A. At the initial period, households are endowed with A₀ and transfer resources intertemporally by borrowing or lending at the real interest rate r_t. A standard intertemporal budget constraint applies:

where Y is labor income, exogenously determined for simplicity ^[6].

Optimizing behavior entails the usual first-order condition:

Equation (4) is a stochastic difference equation that characterizes the intertemporal path of consumption. The expectation term reflects the presence of uncertainty about future variables. Under rational expectations, agents form forecasts using all available information, implying that only unexpected shocks induce deviations in the consumption growth path between t and t + 1. Anticipated events are incorporated into consumption plans and do not generate changes in the expected profile. This distinction between anticipated and unanticipated changes plays a central role in our identification and estimation strategies.

The log-linearization of equation (4) yields equation (5), the basis for our econometric specification:

where:

The left-hand side of equation (5) captures quarterly growth in real consumption per equivalent adult for household i. The error term combines an approximation error and expectational shocks orthogonal to the information set at $t−1$⁠, potentially reflecting macroeconomic or household-level disturbances (Alvarez‐Cuadrado et al., 2016). Consumption growth varies over the life cycle because of sociodemographic changes, such as household composition. The parameter $γ$ denotes the IES, $αi$ captures time-invariant household heterogeneity and $ϵit$ is the composite error term.

Because expectational shocks enter the error term, variables subject to uncertainty, most notably the interest rate, may be endogenous. When households choose consumption at time $t$⁠, the future interest rate is not fully observed, inducing correlation with the error term (Banks et al., 1998). Additional sources of uncertainty include health shocks, income fluctuations and retirement timing, which is voluntary for roughly one-third of households in our sample ^[7].

Under the assumption that the remaining controls are strictly exogenous, consistent estimation requires instrumental variables. To improve efficiency and instrument relevance in first differences, we use Generalized Method of Moments estimators: difference GMM (AB-GMM, Arellano and Bond, 1991) and system GMM (AB-BB-GMM, Arellano and Bover, 1995; Blundell and Bond, 1998). Although typically used in dynamic settings, we apply these estimators in a static framework to strengthen the identification of the interest rate in the 1977–1983 and 1985–1996 panels.

We use several microeconomic data sets from the Spanish National Statistics Institute (INE) that provide detailed information on household consumption, demographics, education and labor market characteristics.

Our main analysis relies on two quarterly rotating panel surveys. The first is EPC, conducted from the second quarter of 1977 to the fourth quarter of 1983. We retain households observed for at least 12 quarters (up to a maximum of 27), yielding a final sample of 29,006 household–quarter observations. The second data set is the ECPF, which tracks households for up to eight consecutive quarters. We use data from the first quarter of 1985 to the fourth quarter of 1996, resulting in an unbalanced panel of 151,513 household–quarter observations after applying standard selection criteria.

The two surveys collect comparable demographic and socioeconomic information, with the ECPF additionally reporting household income. Both are widely regarded as reliable and closely aligned with National Accounts. In the ECPF, total household expenditure represents about 80% of aggregate consumption (around 90% for core categories such as food and housing), while income is underreported and represents approximately 65%. The time profiles of income and expenditure in the microdata closely track their aggregate counterparts, and the degree of underreporting remains stable over time. This stability supports the use of first-differenced variables in our empirical analysis.

More recent data pose methodological challenges. A major redesign of the ECPF in 1998 led to losses in comparability, and in 2006, it was replaced by the EPF, which collects annual data and follows households for at most two years. This short panel dimension makes the EPF unsuitable for our baseline identification strategy, which relies on lagged variables as instruments. We can, however, use the EPF in a robustness exercise comparing the performance of consumption over 1977–1983 and 1985–1996 to more recent patterns. Accordingly, we construct a pseudo-panel using EPF data for 2016–2022. While this approach does not allow for the same econometric techniques as before, it provides complementary evidence that makes our results more robust while helping contextualize the long-run evolution of consumption behavior around retirement.

The Spanish surveys are comparable to widely used international data sets such as the US CEX (Attanasio and Weber, 1995), the Panel Study of Income Dynamics (Bernheim et al., 2001) and the British Family Expenditure Survey (Banks et al., 1998). In several respects, they offer advantages: unlike the CEX and the British survey, the EPC and ECPF provide a longer longitudinal dimension and detailed expenditure information across a broad range of consumption categories. Table 1 reports descriptive statistics for the main variables, and  Appendix 2 provides further details on the surveys.

Household consumption is proxied by the logarithm of expenditure on non-durable goods. As standard in the literature (Banks et al., 1998), we focus on non-durables because durable goods generate intertemporal non-separabilities, are poorly measured in household surveys and account for a smaller share of consumption services (Cutanda et al., 2020). In Spain, non-durables represent roughly 80% of total consumption (Alvarez-Cuadrado et al., 2016). Results are robust to using total expenditure.

To ensure comparability across households, consumption is expressed in per equivalent-adult terms using the OECD-modified equivalence scale and deflated with a household-specific Stone price index ( Appendix 3) ^[8]. Average nominal non-durable expenditure rose from 147,562 pesetas per quarter in 1977–1983 to 544,878 in 1985–1996 and amounted to 23,546 euros per year in 2016–2022.

Nominal interest rates are taken from the Bank of Spain’s short-term policy rate. Household-specific real interest rates are constructed by deflating nominal rates with the Stone price index. Although nominal rates do not vary across households, real rates differ because of heterogeneous inflation. Nominal rates were high during the first two subperiods – reflecting tight monetary policy. It evolved from 13.60% in Q2 1977 to 18.33% in Q4 1983, from 17.6% in Q1 1985 to 8% by the end of 1996 and declined sharply after 2016, remaining near or below zero during most of the EPF period, following the ECB’s expansionary stance in response to low inflation and global uncertainty (e.g. Brexit and US trade policy).

We include a rich set of sociodemographic controls. Household composition is captured by changes in the number of adults, difference in # adults, a key determinant of consumption growth (Attanasio and Browning, 1995). Average household size declined steadily over time, from 3.6 members in 1977–1983 to 2.6 in 2016–2022.

Age of the household head (and its square) controls for life-cycle effects and may proxy time-varying preferences or discounting (Banks et al., 1998). The average age is 53.72 years in the first subperiod, 52.85 in the second and 56.2 in the third, with standard deviations of 12.96, 15.37 and 14.75, respectively. Education is captured by three dummies, primary, secondary and university education (reference category), reflecting substantial educational upgrading over time ^[9]. Labor market status is classified as employed, unemployed or retired, with distributions varying across periods ( Appendix 2).

We capture retirement with a binary indicator variable, dummy for retired, equal to 1 in t if the main income earner in a household is retired, 0 otherwise. In our baseline analysis, we consider as retired individuals self-reporting this status, typically implying receipt of a pension and withdrawal from the labor market.

In the first subperiod, 73.7% of retirees are aged 65 years or older, while 26.3% are early retirees under 65 years. The shares in the second period are 68.4% and 31.6%. These figures reflect the institutional context in Spain. Summary statistics for both groups are comparable, and empirical results using two separate dummies, one for retirees aged 65+ years and another for early retirees, are equivalent ( Appendix 4). This robustness suggests that retirement is largely a planned transition, rather than driven by unforeseen factors. Our data sets do not allow us to identify retirements driven by health shocks, but we conduct robustness checks using alternative definitions of retirement.

We identify the impact of pension reforms with dummy variables. For 1977–1983, we distinguish between nominal pension reforms, which increased pensions without correcting for inflation, and real pension reforms, which raised pensions in real terms. For 1985–1996, we create a pension reform dummy equal to 1 in quarters with real increases in pensions. Additionally, for this period, we construct a continuous variable, average increase in pension, capturing quarterly changes in the average pension at the national level. A similar variable is constructed for 2016–2022. Comparable data are unavailable for the first subperiod.

While this study could have been framed as an evaluation of retirement and pension reforms as “treatments” in a causal inference framework, we opt for a structural approach for several reasons. First, retirement at the statutory age can be treated as exogenous. Second, the timing and nature of pension reforms, implemented heterogeneously across time, complicate a clean identification of causal effects under a treatment-control framework.

We assume that pension reforms were exogenous to household behavior. Although some reforms might have been unanticipated, they were not typically included in the public information set at t. Several arguments support this assumption. First, institutional evidence suggests that reforms during this period were mainly politically and socially motivated, aiming to consolidate the welfare state, rather than to stimulate or contract aggregate demand. Second, even during fiscal consolidation episodes, as in the early 1990s, cuts were implemented in other areas, sparing social expenditures like pensions (Párraga Rodríguez, 2022). Finally, reforms were not tied to macroeconomic indicators like GDP, budgetary rules or life expectancy.

Identifying the relationship between consumption and age is challenging because of the rotating design of our surveys, which prevents long individual-level tracking. To address this limitation, we construct pseudo-cohorts following Deaton (1985), grouping households by the birth year of the household head, as in Banks et al. (1998) and Attanasio and Weber (1995).

For 1977–1983, we consider individuals aged 54–72 years, and for 1985–1996, those aged 54–80 years. Cohorts are defined in four-year birth intervals. Figures 1 and 2 summarize the resulting consumption–age profiles. Panel A reports cohort-specific averages of real non-durable consumption, while Panel B plots marginal age effects from regressions of log consumption on age and cohort dummies.

Across periods and cohorts, consumption generally declines with age, consistent with standard life-cycle predictions (Fernández-Villaverde and Krueger, 2007). An exception arises for younger cohorts in the later period, which display a hump-shaped profile peaking in midlife, likely reflecting earnings dynamics. Notably, no clear discontinuity appears at retirement age, suggesting that retirement effects are embedded in the broader age–consumption gradient. This underscores the difficulty of disentangling age and retirement effects, which are highly correlated in the data.

Several factors may drive the observed decline in consumption with age. Consumption may fall as health deteriorates (Finkelstein et al., 2013), mortality risk rises, discount rates increase or income declines in later life. Because these forces operate simultaneously and are difficult to disentangle, the consumption–age gradient should be interpreted as a composite outcome. Notably, consumption rises slightly after age 70 years among the oldest cohorts, consistent with the positive association between wealth and longevity (Banks et al., 1998), though sample attrition may partly explain this pattern.

Our baseline specification for the periods 1977–1983 and 1985–1996 is given by:

Because the data consist of household panels observed over multiple quarters (unbalanced panels), in this section we exploit within-household variation without constructing pseudo-panels.

We begin by estimating a Euler equation including standard demographic controls: the change in the number of adults in the household and the age of the household head, interpreted as a proxy for a time-varying discount factor. Age squared is added to allow for nonlinearities, consistent with the patterns in Figures 1 and 2. The specification closely follows Banks et al. (1998, p. 775), with the key addition of a retirement dummy. Educational attainment is controlled for using two differenced dummies (dif. primary and dif. secondary education, with university education as the reference). Quarterly dummies capture seasonality, with the fourth quarter as the baseline. Robust standard errors are clustered at the household level.

Table 2 reports the estimation results. Panel A corresponds to 1977–1983. Column (1) presents random-effects GLS estimates ^[10]. The interest rate enters with a positive but insignificant coefficient. Changes in household size are positively and significantly associated with consumption growth, while age of the household head enters negatively and significantly, consistent with its interpretation as a discount-factor proxy (Banks et al., 1998) ^[11]. The retirement dummy is negative but insignificant, suggesting no consumption discontinuity at retirement conditional on expectations, in line with the LCPI.

Given the potential endogeneity of the interest rate, Column (2) reports two-stage least squares estimates, instrumenting the interest rate with four lags. The results for the controls are unchanged, while the interest rate becomes negative and statistically significant. The retirement dummy remains insignificant.

Columns (3) and (4) report two-step GMM estimates using AB GMM and AB–BB GMM estimators, treating the interest rate as endogenous and using the same lag structure as instruments ^[12].AB GMM yields a positive and significant interest rate coefficient, implying an IES of 0.30 (relative risk aversion of 3.3). The AB–BB GMM estimate implies a lower but more precisely estimated IES of 0.149. In both specifications, the retirement dummy remains statistically insignificant.

Panel B of Table 2 reports results for 1985–1996, using the same specification and instruments to facilitate comparability. The interest rate coefficient is positive and significant across all models, with AB–BB GMM again implying a relative risk aversion coefficient close to 3.3. In contrast to the earlier period, the retirement dummy is negative and significant in all specifications. According to the GMM estimates, retirement reduces consumption growth by between 1.1% and 4.9%, consistent with prior evidence of a retirement consumption drop.

The role of demographic variables differs across periods. In 1985–1996, age enters with a small positive coefficient, significant only in the AB GMM specification, while changes in household size are negative and significant throughout. We interpret the sign change in the age coefficient as reflecting evolving preference parameters over time. Concurrently, average household size declined markedly, from 3.63 members in 1977–1983 to 2.80 in 1985–1996, consistent with broader demographic trends such as declining fertility, increased female labor force participation and changing household structures.

We extend the baseline specifications by adding controls for pension reforms. For 1977–1983, we distinguish between nominal and real pension reforms, while for 1985–1996 we include a single dummy (pension reform) capturing real pension increases. All GMM estimations continue to instrument the interest rate using four lags ^[13].

Table 3 reports the results. In the first subperiod, the interest rate is generally insignificant, except in Column (2), where it enters with a negative and significant coefficient. This lack of robustness likely reflects correlation with unobserved heterogeneity, such as time-varying impatience. Age enters negatively and age squared positively under GLS and 2SLS, consistent with nonlinear ageing effects. Educational attainment remains insignificant, reflecting limited within-household variation.

The retirement dummy is negative and significant across all specifications, implying a consumption drop between 9.4% and 14.8%. Both nominal and real pension reform dummies are positive and significant, indicating that pension increases raised consumption and partially offset the retirement-induced decline. Real reforms exert a stronger effect than nominal ones.

Relative to Table 2, the inclusion of pension reform dummies alters both the magnitude and significance of the retirement coefficient, suggesting that baseline estimates conflated two opposing effects: a negative retirement shock and a positive policy-induced income shock. Table 3 disentangles these mechanisms. These findings are consistent with Párraga Rodríguez (2022), who documents a high marginal propensity to consume out of pension increases, and provide further evidence against the LCPI using a structural approach.

Panel B presents results for 1985–1996. The interest rate remains positive and significant, confirming the robustness of earlier estimates. Retirement is again associated with a significant consumption decline of around 5%, smaller than in the earlier period. The pension reform dummy is positive and significant in most specifications and, in the AB GMM model, more than offsets the retirement effect. Controlling for reforms raises the AB–BB GMM estimate of the IES to 0.572, compared to 0.302 in Table 2. This estimate aligns with those reported by Attanasio and Weber (1995) for the USA, Banks et al. (1998) for the UK and Cutanda et al. (2020) for Spain, underscoring the importance of policy-driven income shocks for consumption dynamics.

Taken together, Tables 2 and 3 reveal a robust negative association between retirement and consumption in both periods. The estimated drop is larger in 1977–1983 (approximately 10%–13.5%) than in 1985–1996 (1%–5%) and is partially mitigated by pension reforms, particularly in the earlier period. These results challenge the LCPI prediction of smooth consumption and policy irrelevance at retirement.

The evolving institutional context in Spain helps explain these patterns. During 1977–1983, financial markets were underdeveloped, limiting households’ ability to smooth consumption intertemporally and weakening the link between interest rates and consumption. By contrast, the liberalization and financial deepening of 1985–1996 expanded saving opportunities, consistent with the stronger interest rate effects observed.

Additional mechanisms may also contribute to the consumption drop at retirement. Retirement may heighten perceived health risks, triggering precautionary behavior or reveal adverse information about actual pension entitlements relative to prior expectations (Banks et al., 1998). These channels are particularly relevant in Spain, where frequent reforms, high inflation and institutional opacity complicated retirement planning.  Appendix 4 explores these issues and presents further robustness checks.

A further explanation is the arrival of adverse information about post-retirement income. Individuals may overestimate future pension entitlements while working and revise consumption downward once actual benefits are realized (Banks et al., 1998). This channel is particularly relevant in Spain, where the pension system was evolving rapidly and frequent reforms increased complexity. Opaque rules, multiple regimes, high inflation and limited financial literacy likely hindered the formation of accurate income expectations.

Political uncertainty and limited institutional credibility in the early years of Spanish democracy may have further amplified perceived risks. The prospect of abrupt fiscal adjustments, especially around government changes or in the run-up to Maastricht convergence, could have discouraged forward-looking consumption, particularly among more vulnerable households.

We assess the robustness of our results to alternative specifications in  Appendix 4.

This paper revisits the retirement–consumption puzzle using rich household data for Spain and finds systematic departures from the predictions of the LCPI. We document a significant decline in consumption growth at retirement that progressively weakens over time: around 10%–13.5% in 1977–1983, 4%–5% in 1985–1996 and about 2% in 2016–2022. Pension reforms are positively associated with consumption growth in the first two subperiods, indicating that current income affects consumption decisions, particularly in contexts of institutional uncertainty.

We find no evidence that the consumption decline is driven by reductions in work-related expenses or changes in health expenditure. Health spending represents a small share of non-durable consumption and remains stable at retirement, likely reflecting Spain’s extensive public health coverage.

Instead, the evidence points to uncertainty, especially regarding future pension income in a pay-as-you-go system, as the main driver of the observed consumption drop. Limited financial literacy and the complexity of pension rules likely hinder accurate expectation formation, particularly during periods of weaker institutional credibility and macroeconomic instability. This interpretation is consistent with the weaker link between consumption growth and interest rates in the early period and its strengthening over time as financial markets deepened.

From a policy perspective, these findings highlight the importance of institutional clarity. Greater transparency in pension rules, improved financial education and policies that facilitate saving and decumulation in retirement could reduce uncertainty and promote smoother consumption paths, with potentially beneficial macroeconomic effects.

This study has limitations. The analysis for 2016–2022 relies on pseudo-panels, which reduces statistical power and limits identification. Future research could exploit alternative data sources and causal strategies, such as difference-in-differences or regression discontinuity designs. Further work should also focus on measuring pension complexity and financial development more directly, as well as exploring heterogeneity by income, education and household composition to better inform targeted policy interventions.

It defined two pension categories: Contributory, linked to workers’ contributions over the last period of professional life; and Non-contributory, for retirees with income below a threshold not entitled to i.

The statutory retirement age was 65 years. Early retirement was possible under specific conditions. Retirement implied withdrawal from the labor market. The law established minimum and maximum pensions and annual adjustments.

Contributory monthly pension for a worker retiring at 65 years:

α_n= replacement rate, depending on the years of contribution; and

BR_t= benefit base, weighted average of monthly earnings over a reference period.

EPC= quarterly survey of households, 2,000 per wave, 5% randomly replaced each quarter and 130 goods and services.

ECPF= superseded EPC, quarterly, wider coverage, 3,200 households per quarter and 12.5% renewed per wave.

EPF= replaced ECPF in 2006, annually, 24,000 households per year staying two consecutive years, 2,000 interviewed each month and 50% renewed monthly.

Constructed as Stone indexes, with average shares of expenditure as household-specific weights.

Let $cjkh$ be consumption of household h in category j in quarter k, where h participates during K quarters. j indexes goods. Let $TCkh$ be total consumption of h in k. For household h, the share of good j in k:

And the average share overtime for j:

Increase in Stone price in t:

where p_jt is the price for good j in t and $Pth$ the Stone aggregator for h in t.

We summarize the results for several robustness tests, using AB-BB GMM and instrumenting the real interest rate with lags 1–4 (unless otherwise specified).

Work-related expenditures and restricting the definition of retirees

To assess if the consumption dip at retirement is because of the decline in work-related expenses, we re-estimate equations (5) using the first difference of (log) work-related expenditures as dependent variable. Results indicate no significant relationship between retirement, pension reforms and work-related expenditures (Table A2, first column). For 1985–1996, the coefficient of the interest rate is large and significant (0.789), implying a higher IES, consistent with the notion that individuals postpone non-essential purchases when real returns rise.

To test if the retirement effect is driven by early retirees, we re-estimate the model using a more restrictive definition of retirement (household head aged 65 years or older, retired). Results (Table A2, Columns 2 and 3) are virtually identical to those using the broader definition, suggesting that retirement is largely anticipated and incorporated into household planning.

Enriched specifications.

We decompose the impact of retirement in two components. The “impulse effect” – dummy, retiring in quarter (impulse) – captures the immediate consumption response in the quarter of retirement, while the “step effect” – dummy for being retired (step) – represents the persistent but attenuated influence of being retired in subsequent periods. Both effects are negative and significant (Table A3) with the short-run impulse stronger than the step, consistent with a sudden adjustment followed by a gradual adaptation.

We explore the determinants of early retirement by instrumenting these two retirement dummies with lags 1–4 of total household income, proxying for financial conditions affecting exit timing from the labor market. Pension reforms are controlled for with a new variable measuring the yearly average increase in pensions (Ministerio de Inclusión, Seguridad Social y Migraciones, 2023). Results (Table A3, Column 2) are similar to the baseline, with the new measure of pension increases positive and significant. Instrumenting the retirement dummies with lagged income percentiles (Column 3) yields consistent findings.

We incorporate a macroeconomic control, the quarterly real GDP growth rate, instrumented by lags 1–4 to account for potential endogeneity through correlation with the interest rate. It is positive and significant (Table A3, Column 4), suggesting that broader economic conditions influence household consumption.

Results for 2016–2022.

As an additional robustness check and to connect the analysis to current policy debates, we construct a pseudo-panel from the EPF for 2016–2022 following Deaton (1985). Because EPF interviews households at most twice, internal instruments cannot be exploited once first differences are taken. We therefore group households into 10-year birth cohorts of household heads, yielding a monthly pseudo-panel of 588 observations (7 cohorts × 7 years × 12 months). Expenditures are deflated and equivalized as in earlier sections.

We re-estimate equations (5) on cohort means using 2SLS to address interest rate endogeneity, alongside random effects estimates for comparison. Age controls are replaced by mortality risk as a nonlinear proxy. Results are reported in Table A4.

The retirement dummy remains negative and statistically significant, with an estimated consumption decline of about 1.9%–2%. This magnitude is consistent with earlier findings and confirms a declining retirement-consumption gap over time: 10%–13.5% in 1977–1983, 4%–5% in 1985–1996 and around 2% in 2016–2022. This pattern is consistent with Spain’s increased financial development and macroeconomic stability, which likely reduced uncertainty and improved retirement planning. Behavioral and structural factors, such as better health, expanded leisure opportunities and the growth of the silver economy, may have further attenuated the consumption drop. The small and imprecise interest rate effect likely reflects the prolonged period of near-zero and negative rates.

The consumption contraction observed in 2020 because of COVID-19 could potentially bias the estimated retirement-related consumption drop. However, this effect is largely absorbed by the inclusion of year dummies for 2020 and 2021. In our data, non-durable consumption fell by 6.98% between 2019 and 2020 and rebounded by 7.03% in 2021, implying that the pandemic shock was largely transitory. Moreover, although the point estimates of the 2020 and 2021 dummies differ, their 95% confidence intervals substantially overlap, suggesting no statistically meaningful asymmetry between the two effects.

Health expenditures.

We also examine whether changes in health-related expenditure contribute to the retirement consumption decline. Health spending, covering insurance, pharmaceuticals and medical devices, accounts for only about 2% of non-durable consumption, with similar shares among retirees and non-retirees. Estimating equations (5) using real health expenditure per equivalent adult as the dependent variable yields no significant effects of retirement or age (Table A5), including specifications restricted to retired households and those using mortality risk instead of age. These results indicate that health expenditures do not drive the observed consumption drop, likely reflecting Spain’s extensive public health coverage.

Housing and interactions.

We test whether home ownership affects consumption by including a housing dummy. Results (Table A6) reveal no significant relationship, suggesting that housing wealth does not translate into higher liquidity or expenditure. Interactions between retirement, pension reforms and age yield similar conclusions.

Additional robustness exercises produce results consistent with the baseline and are available upon request.