Beyond the headline: How personal exposure to inflation shapes the financial choices of households

The paper asks whether individual consumption spending responds to the inflation a household is personally exposed to, over and above the headline rate, and whether financial constraints modulate that response. The authors state that their “analysis is the first to link the inflation experienced at the group level to the dynamics of consumption spending and to explore how the heterogeneous financial constraints faced by households modulate this link.” The identification problem they name is twofold: for most countries, high and variable inflation belongs to past periods when granular data were scarce, and even today current inflation “is, or at least can be, endogenous to current consumption spending in most setups, or both inflation and consumption spending are affected by common drivers like monetary or fiscal policy.”

Estonia is chosen because it combines unusually large variability in inflation over the 2000s with a small-open-economy structure in which inflation is largely driven by shocks from abroad, which the authors argue mitigates the reverse-causality problem. Estonia also operated a currency board with the euro during the sample, so nominal interest rates followed rates set abroad. The authors note that the standard channel by which higher inflation triggers an expected policy-rate rise “does not operate in a small open economy with fixed exchange rates when uncovered interest parity holds, as domestic interest rates follow global rates and the central bank has limited leeway to adjust interest rates in response to domestic inflation.” Two identifying assumptions are stated explicitly: first, that inflation in Estonia is not driven by short-term domestic demand; second, that there are no significant unobserved variables correlated with personal inflation exposure and with spending that vary over time in a way the unit fixed effects do not difference away.

An anonymised quarterly database from a major commercial bank in Estonia, covering inflows to and outflows from individuals’ checking accounts and end-of-quarter balances of checking and savings accounts, securities and debt, running from Q4 2004 to Q4 2011. The full dataset covers over 100,000 individuals, roughly 12% of Estonia’s working-age population at the time; the final estimation sample is 89,507 individuals, with the time dimension varying between 24 and 29 quarters. Banking concentration in Estonia was very high – the three largest banks held 96% of banking-sector assets – and the sample is drawn from the regular customer base covering regular income and all transactions, so the authors assume customers treat the bank as their main bank and that the data capture the complete picture of their financial transactions and financial assets. The main limitation is that the account data contain no product-level spending, so expenditure shares must be imputed.

Group-level inflation is a weighted average of category-specific inflation rates, with the weights coming from the Estonian Household Budget Survey’s 12 consumption categories (waves 2005-2007 and 2010-2011), imputed onto the account data from income decile, ten age groups, a dummy for residence in Tallinn and a gender dummy – effectively 400 groups. The authors are careful about terminology: because weights are group-specific rather than truly individual, the constructed index “should be interpreted as a group-specific price index,” they use “personal” and “group-specific” interchangeably, and they note that the regressor “strictly captures ‘inflation exposure’.” The cost is compression: imputation “compresses the standard deviations in the imputed shares,” and because only between-category price variation is used, dispersion is smaller than in scanner-data studies that exploit within-category variation (Kaplan and Schulhofer-Wohl 2017). Measurement error of this kind produces attenuation bias, which “would work against us finding any significant results by biasing the coefficients towards zero” – hence the authors’ repeated characterisation of their estimates as lower bounds.

Mean quarterly personal inflation averages 1.2% over the period, with a within standard deviation of 1.11 and a between standard deviation of 0.07; the largest gap across groups is two percentage points in 2008Q4, and the average gap over the period is 0.72 percentage points. Quarterly personal inflation peaks at 4% in 2008Q1 and troughs at -2.71% in 2009Q2; in annual terms the highest group rate is 13.8% and the lowest -3.7%. Dispersion is larger when aggregate price changes are large, so, in the authors’ words, heterogeneity “matters more during turbulent periods.” Low-income and elderly groups faced higher inflation when prices rose but lower inflation during the 2009Q2 deflation. Persistence is low: the autocorrelation coefficient is 0.14 for quarterly and -0.25 for yearly personal inflation, so the set of households experiencing relatively high inflation rotates over time.

In the baseline two-way fixed effects specification the coefficient on quarterly personal inflation is 0.014 (standard error 0.003), i.e. a one-percentage-point increase in mean quarterly personal inflation raises real consumption spending by 1.4%. Because individual and time fixed effects are included, headline inflation is absorbed by design, and the coefficient is the response to personally experienced price changes as a deviation from the headline rate. Allowing a quadratic term gives a positive and significant coefficient of 0.005, indicating stronger responses at higher levels of inflation – which the authors read as consistent with Cavallo et al. (2017) on weaker priors about inflation in low-inflation environments. Replacing two-way fixed effects with interactive fixed effects, which impose a factor structure on the errors to absorb time-varying unobservables, gives 1.5% under the Vogt et al. (2022) optimal-factor approach and 1.4%, 1.4% and 1.1% under Bai (2009) with one, two and three factors respectively. Standard errors are clustered at the cell level defined by the time-invariant imputation variables.

No – 1.4% is the direct, partial-equilibrium intertemporal-substitution response; the total effect the authors report, after netting the negative real-income and real-wealth effects, is “positive and close to 1%.” They document the arithmetic: a 1% fall in real income lowers spending by 0.48% and a 1% fall in real assets lowers it by 0.11%, and one percentage point of extra inflation reduces real income and real assets by roughly 1% each, so the net across the three channels stays positive and near 1%. They also flag what is not in the estimate: general-equilibrium feedback of the kind in Kaplan et al. (2018), where lower real rates raise consumption and then labour income, is absent, because the model “does not estimate indirect effects stemming from broader economic dynamics, such as employment impacts.”

Partly by running down term deposits and partly by borrowing. A one-percentage-point rise in quarterly personal inflation is associated with a fall of roughly 12% in term-deposit balances (coefficient -0.119, standard error 0.020), while checking-account balances show no significant relation (0.008, standard error 0.010) – individuals appear to keep a buffer in the checking account and draw on term deposits instead. On the borrowing side, the estimated odds ratios are 1.142 for adding a consumer loan and 1.055 for adding an overdraft, both significant, while the odds ratio for adding a housing loan (0.949) is not. Stock holdings respond positively (0.093, significant); the authors flag an unresolved tension here, since Estonian stock-index returns correlate negatively rather than positively with headline inflation, and leave the resolution to further research.

Liquidity amplifies the response and debt dampens it. Interacting personal inflation with the lagged ratio of liquid assets to annual income yields positive and significant coefficients whether liquidity is measured in checking accounts, term deposits, or the two together, so individuals with larger liquid reserves respond more strongly; the authors read the pattern as liquidity constraints preventing others from financing the front-loading of spending. A triple interaction with a dummy for quarters in which personal inflation fell is negative and significant (-0.022), implying liquidity matters much more when inflation is rising than when it is falling – as expected, since the incentive to bring spending forward is weaker when inflation is falling. Interactions with debt ratios are negative and significant for total debt (-0.004), housing loans (-0.003) and consumer loans (-0.048), and the interaction with the debt servicing ratio is likewise negative (-0.018). The authors conclude that a high level of debt “can act as a binding constraint on the ability to bring consumption forward, and that this mechanism prevails over the positive debt depreciation effect.” In the sample, individuals deplete their liquid resources by the end of the pay period in 34% of individual-quarter observations and hold some form of credit in 46% of cases.

Food and transport, not housing. Re-estimating the baseline with category-level inflation for the three largest expenditure categories gives positive and statistically significant coefficients of 0.009 for food and 0.063 for transport, while the coefficient for housing (0.006) is not significant. The authors note that fuel prices affect producers’ and retailers’ transport costs as well as households’ own, so spillovers into other prices may explain why the transport response is the stronger one.

Front-loading happens within the quarter and the following one, and then reverses. Adding lags of personal inflation to the baseline gives a still-positive coefficient of 0.021 on inflation experienced one quarter earlier and a negative coefficient of -0.027 on inflation experienced two quarters earlier, with similar results from a distributed-lag specification. The authors describe this as “a rather slow response in spending,” and the sign reversal is what intertemporal substitution implies: an increase in spending must be followed by a decline.

It argues expectations are the most plausible explanation but is explicit that the core estimates do not test channels directly. The authors write that “the most plausible explanation for this finding seems to be the expectations channel, although our core estimates do not directly test for different channels,” and state their conjecture that individuals experiencing higher inflation raise their expectations of future inflation proportionately more and therefore raise current spending – supported with Estonian survey data in an appendix table. They deliberately bypass measured expectations, specifying spending directly on recent personal inflation exposure so as “to remain agnostic about the exact process of expectation formation.” They also note that the sign was not a foregone conclusion: households that regard rent, food or transport spending as unshiftable could instead have cut consumption and raised precautionary saving, and the existing literature on expectations and spending contains mixed and negative findings as well as positive ones.

Three implications, all conditional on the fact that current mandates target the mean rather than the tails of the inflation distribution. First, larger spending increases by households already more exposed to inflation may reinforce demand for the goods whose prices are rising fastest and so reinforce inflation patterns, which the authors say makes central-bank credibility in fighting inflation “all the more important” for breaking or moderating the link from recent inflation through expectations to current spending. Second, because responses are moderated by liquidity and borrowing constraints, households’ ability to smooth consumption – and hence the pass-through of monetary policy – depends on the degree of those constraints. Third, because food and transport prices do much of the work in household spending decisions while central banks target core inflation excluding food and energy, the authors suggest that giving non-core categories closer consideration “could contribute to a more responsive and efficient policy approach.” They add that group-specific inflation of the kind they construct can be computed and monitored by central banks, so focusing on a limited number of population groups is a practical option.