The effect of price changes experienced by consumers when doing purchases on their inflation forecasts

The effect of price changes experienced by consumers

when doing purchases on their inflation forecasts

Nils Brouwer

June 10, 2019

Abstract

This paper examines the relationship between the price changes experienced by households when doing purchases (household-specific inflation) and their 12-month ahead inflation forecast, a relation which is still little investigated. Gaining a better understanding of the impact of household-specific inflation on the price-forecasts of consumers provides new insights on how individuals form their expectations and helps to explain why it differs across individuals of different gender, age, and income. The household-specific inflation is calculated using micro-expenditure data and matched to a survey containing price expectations based on observable characteristics (age, gender, and income). The results show that: the household-specific inflation has an impact on the inflation expectations of consumers (I) and agents overweigh the most extreme-experienced inflation of the past six months and gas prices when making short-term price forecasts (II).

1

Introduction

Since the work of Lucas and Rapping (1970), the representative agent model has been one of the most predominant means of studying economic behavior.1 The Global Financial Crisis (GFC) of 2008, however, has revealed the need to rethink how academics and policymakers should model the aggregated economy. The disregard of heterogeneity among agents in economic models has led to the failure of economists to anticipate on the GFC and made it hard to predict the outcome of the policy conducted by central banks (Colander et al. (2009)).

One of the fields in which it is crucial to allow for heterogeneity is the expectation formation of individuals. Survey data shows substantial disagreement in the expectations of future inflation among consumers (Mankiw et al. (2003)) and these differences have a significant impact on the dynamics of the economy at an aggregated level. Heterogeneity in consumers’ inflation expectation can generate consumption inequality (Di Bartolomeo et al. (2016)), can lead to the creation of bubbles (Branch and McGough (2018)), and impacts the economy’s vulnerability to shocks (Badarinza and Buchmann (2011)). Furthermore, even before the start of the GFC the Federal Reserve has repeatedly stressed the importance of understanding the heterogeneity in inflation expectations for the transmission of monetary policy (Bernanke (2007)), a channel which became even more critical for central banks since

nominal interest rates have moved towards the effective lower bound in the past ten years (Bernanke et al. (2019); Duca et al. (2018)).

Despite the agreement on the need to allow for heterogeneity, there is still little con-sensus on the best way forward. Heterogeneity implies that economists have to reject FIRE (full-information rational expectations hypothesis) for which there are several options. Economists have used various methods to reject FIRE, while simultaneously holding their models traceable, of which: heuristics, information rigidities, different-forecasting models and learning models (Coibion et al. (2018)) are the most notable. However, since there is no conclusive evidence which model is best, the discussion of how to model the variaty in expectations is still an ongoing debate among economists.

So far, there are only a few empirical studies conducted which can explain the hetero-geneity in the inflation forecasts of agents. For example, experimental research shows that individuals use different models (Pfajfar and ˇZakelj (2014)) and use different information (Armantier et al. (2016)) to estimate expectations. However, applying these mechanisms on survey data reveals that there is still much work to do: the predictive power of these models is low, and proper identification of why agents differ remains difficult.

This paper aims to provide new explanations for the question of how individuals form their price forecasts. As answering this question is critical to explain the heterogeneity found in inflation forecasts among individuals. It will do so by investigating the effect of inflation rates experienced by households when purchasing goods (household-specific inflation) on the formation of expectations. The household-specific inflation rate can vary across households because they buy different bundles of products (Michael (1979)) and prices differ between regions. Until now, most papers assume this effect to be small (e.g., Bruine de Bruin et al. (2010)) because they assume that in the long run, the household-specific inflation rate does not deviate substantially from the aggregated inflation rate (Hobijn et al. (2009)).

However, recent research shows that the volatility of inflation rates is substantially dif-ferent for various socio-demographic groups and the cumulative differences in inflation rates across income groups, in particular, are substantial (Kaplan and Schulhofer-Wohl (2017)). Therefore the interest of researchers in this topic has reopened recently. In a recent working paper, which is closely related to this paper, D’acunto et al. (2019) show that household-specific inflation is one of the most important drivers of inflation expectations. They test this using a two-time survey in (2014 and 2016) asking 40.000-60.000 US consumers about their inflation expectations and recording their purchases in the period 12 months before both interviews. This paper will extend their research by looking at the same effect but uses a different dataset which covers a more extended period (14 years) and by building further on their effort of explanation how household-specific inflation influences expectations.

age) and tested to see whether it has predictive power.

The main take away of this paper is that household specific inflation has a significant (positive) effect on the inflation expectations of households. An increase of 1 percentage point in the experienced inflation increases the forecast by 0.03 to 0.06 percentage point. This effect, as the household-specific inflation differs among agents, can explain part of the heterogeneity of consumers’ price forecasts. Furthermore, this paper finds three other in-teresting phenomena. There is no evidence that the volatility in household-specific inflation increases the likelihood of updating their expectations (I), indicating that it does not play a role in models using information rigidities. Individuals use household-specific inflation even when they read news regarding inflation (II). This implies, in contrast to the findings of D’acunto et al. (2019), that the effect of household-specific inflation is not crowded out by the use of other sources of information. Also, this paper finds that individuals use household-specific inflation in their heuristics. Changes in gas prices, the most extreme value of the household-specific inflation and goods which are bought frequently are over-weighted (III). The rest of this paper is structured as follows. The first section will give an overview of the reasons why current economists often use bounded rationality and how household-specific inflation might influence expectations. The second section will elaborate on the data used in this paper to calculate household-specific inflation, the data on inflation expectations, and how these two are matched. The third section shows the results for the effect of household-specific inflation in general and the three channels. The fourth section concludes.

2

Literature review

2.1

A brief history of modeling expectations

2.1.1 The rejection of FIRE

Until the beginning of the twentieth century, most economists used the full-information rational expectations hypothesis (FIRE), first adopted by Muth (1961), when looking at the expectations formation of consumers. FIRE makes two core assumptions about the behavior of agents: agents have full information, and they do not make persistent errors over time. If we accept both these assumptions than agents have homogeneous expectations by definition (Demery and Duck (2007)); a prediction which is clearly false in the case of inflation expectations. As there is ample empirical evidence which shows the existence of persistent and substantial heterogeneity in the expectations of consumers in for example Sweden (Dr¨ager (2015)), Germany (Menz and Poppitz (2013)) and the US (Mankiw et al. (2003)). Economists who want to explain the heterogeneity in these surveys, therefore, have to reject FIRE for which they have three options: reject full-information (I), reject rationality of agents (II) or reject both (III).

In the last two decades, the first option became the standard approach in research about the formation of expectations. There are two main reasons why economists choose this op-tion. First, researchers want to avoid the rejection of rationality since rejecting rationality allows agents to act ’randomly’ which makes it hard to make any predictions. Furthermore, researchers found, in an experimental setting,that individuals do not use all available infor-mation (e.g. Armantier et al. (2016)). The practice of rejecting full-inforinfor-mation will also be used in this paper.

or face constraints when modeling the best course of action (but are still rational given these constraints).2 The models which use bounded-rationality outperform the assumption of FIRE because it allows the information set of agents, and the model to process this information, to differ across agents which can create the discord found in surveys.

2.1.2 Incorporating bounded-rationality in expectation formation models Bounded rationality, however, also has significant drawbacks. Sims (1980) shows the prob-lem of the ‘wilderness’ of bounded rationality. Since there are too many degrees of freedom and too many free parameters, there are simply too many ways of modeling non-rational behavior. This ‘wilderness’ of bounded rationality, seems particularly problematic when one allows individuals to have heterogeneous expectations (Hommes (2011)).

Contemporary economists avoid such ’wilderness’ by making assumptions about the form of bounded-rationality when specifying an expectation formation model. There are four popular approaches used by economists to model the expectation formation of boundedly rational agents, which are summarized by Coibion et al. (2018): information rigidities, heuristics, the different-forecasting/information approach and learning models. The first three of these models will be further explained in the next section, where the effect of household-specific inflation in these models is further explained. The last approach, learning models, will not be used in this paper as the preferred data (panel-data) is not available and the, for this purpose, lower-quality data (cross-sectional) requires an assumptions which decreases the variance among the experienced inflation levels of households substantially. This approach, popularized by Malmendier and Nagel (2015), is mostly used to look at how expectations change over time by examining how an individual changes his expectations when confronted with an information shock. This approach requires either a dataset which tracks the same individuals over time or an often repeated cross-sectional survey to create pseudo-panels. A data-set which asks the same individuals over a longer-time period about their inflation expectations is not available as far as the author of this paper is aware. Pseudo-panels can only be created based on characteristics which are time-invariant such as birth-cohort (Verbeek (2008)). This means that we cannot look at the differences in inflation which might arise because of time-variant variables such as income, a factor which according to the literature is the most important determinant of differences in inflation among households (see for example Flower and Wales (2014); Kalwij et al. (2018); Kaplan and Schulhofer-Wohl (2017)).

2.2

The role of household-specific inflation on expectations

Up to now, there is only little empirical work on the effect of household specific-inflation on expectations. Most papers assume that the impact will be small because research at the beginning of this century found that household-specific inflation, in the long run, did not deviate substantially from the general inflation level (Hobijn et al. (2009)). Therefore, researchers focused on other mechanisms which can explain the heterogeneity found in in-flation expectations. For example: Pfajfar and Santoro (2013) found that the amount of news coverage about inflation increases the accuracy of inflation expectations, but that the effects of news coverage differ per demographic group, Dr¨ager and Lamla (2012) show that the trust in a government can explain differences between expectations and Souleles (2004) showed that inflation forecasts are positively related to the sentiment of individuals towards

other economic indicators. However, despite all these various models, the predictive power of explaining the heterogeneity remains low.

In more recent research, Kaplan and Schulhofer-Wohl (2017) find evidence of vast dif-ferences in the household-specific inflation in the US between 2000 and 2013. These new findings reopen the question of whether household-specific inflation plays a role in the ex-pectation formation process of households.

The only research so far, of which the author of this paper is aware, which uses household specific inflation, is the paper of Pfajfar and Santoro (2009) and the working paper of D’acunto et al. (2019). Pfajfar and Santoro (2009) show that, using an American consumer survey, the inflation expectations of low-income individuals are closer to their experienced inflation when purchasing goods than the aggregated level of inflation. However, they do not go into further detail to explain as to why this might be the case.

D’acunto et al. (2019) recorded the goods purchased by a sample of 40.000-60.000 US consumers during two 12-months periods using the scanner-data of the Nielsen scanner data. The Nielsen scanner data tracks the purchases made by a group of individuals over time by asking them to scan the barcode of all their purchases and report the price they paid for this good. By using the method of Kaplan and Schulhofer-Wohl (2017), the authors create a household-specific inflation based on this data in which the consumption bundles and the prices paid for a particular good vary among individuals. They surveyed all participants at the end of each 12-month period asking a set of questions regarding their expectations of future inflation. This way, the authors could directly test whether the price changes faced by consumers during these 12-months periods had an impact on their price expectations. Their results indicate that the household-specific inflation is a key determined in explaining inflation forecasts and that this effect is mainly driven by limited attention of agents towards more representative information about overall information (such as core-inflation rates or price forecasts by professionals).

This paper distinguishes three channels of how household-specific inflation might in-fluence the expectations of consumers. These channels are not only meant to show ’how’ household specific inflation influences expectations but also intends to link these effects to a particular approach of bounded rationality, which was introduced in the previous section. This way, we can see how the household specific inflation fits into contemporary theories regarding bounded-rationality and the relevance of this paper for future research, using one of these approaches becomes clear. These channels are empirically validated in section 4.

2.2.1 Salient-price channel

The salient-price channel assumes that agents will use heuristics to determine their inflation expectations. Agents will not rely on all available information but will use ”rules-of-thumb” to calculate their inflation forecasts. Two of the most researched ”rules-of-thumb” to explain individuals’ inflation forecasts are: agents will only rely on price information of goods they frequently purchase (I), and they will only recall the most extreme changes in inflation when calculating their price forecasts.(II).

which are purchased sporadically when calculating the inflation perceptions of individuals. A possible explanation for this finding is that individuals are better able to see price changes in these goods because the time between each purchase is smaller. In the case of household-specific inflation, this means that individuals will overweigh the price changes of goods which they buy more frequently.

D’acunto et al. (2019) find evidence for this effect by comparing the effect of a household-specific inflation based on a ’frequency-basket’ to an ’expenditure-shares-basket’. When both of these measures are used, the coefficient of the ’expenditure-share-basket’ inflation becomes non-significant. This paper will try to replicate this result by calculating, and testing the effect of, both measures of household-specific inflation.

Hypothesis 1 The household-specific inflation based on ’frequency-baskets’ outperforms the measure using ’expenditure-baskets’.

Individuals find it easier to recall events which are atypical or extreme (Brown and Kulik (1977)). Morewedge et al. (2005) use this insight from psychology and find in an experimen-tal setting that agents rely on atypical events when making forecasts regarding economic variables. These authors argue that this might explain why the estimates of economic vari-ables made by consumers tend to be consistently above the actual value. Individuals use highly available but unrepresentative memories of the past (big shocks in prices) when mak-ing their forecasts. This means that households will use periods of unusually large changes in experienced prices more than periods where prices are more stable if asked about their inflation forecasts.

Similarly, economists have argued that the prices experienced ’at the pump’ (changes in the price of gas) are overweighed by individuals when making forecasts (Bruine de Bruin et al. (2011); Binder (2018)). Gas prices are volatile, purchased frequently and visible (for example by signs at the road near gas pumps), which makes it easier to recall these price changes when making inflation forecasts.

Hypothesis 2 Households overweigh periods of relative extreme-household specific inflation when making inflation forecasts.

Hypothesis 3 Households overweigh the experienced price changes of gas when making inflation forecasts.

2.2.2 Updating channel

Mankiw and Reis (2002) introduce the updating channel in their sticky-information model (information rigidities), a model which is derived from the sticky-price model of Calvo (1983). Agents do not update their information set or recalculate their expectations in real-time because there are costs associated with this process. An agent will only gather new information or update their calculations if the expected rewards are higher than the expected costs. Once the expected rewards are higher than the expected costs, the agent will gather all relevant information, and the prediction of an agent at that time is again FIRE. This mechanism implies that for every new period, only a fraction (λ) of the population will update/recalculate their information set. This theory generates heterogeneity in expecta-tions because different segments of the population will have updated their expectaexpecta-tions at different points in time.

of 500 US consumers what their price expectations are in the next 12 months.3 They do not derive the fraction (λ) endogenously but choose it such that it maximizes the correlation between their estimated model and the data.

Several papers have tried to find endogenous explanations for the proportion of individ-uals who will update after each new period. Most of these explanations look at how the rewards and costs of updating are determined. Two often used factors are the amount of news and the volatility of economic variables. The amount of news decreases the cost of acquiring more information since it is more readily available. Moreover, an increase in the volatility of economic variables increases the reward since a larger degree of volatility might signal a shock, and updating faster will lead to a decrease in an agents exposure to the risks caused by this shock.

Dr¨ager and Lamla (2017) test these two effects empirically by looking at the probability that an agent updates his/her expectations. They use the rotating-panel component of the Survey of Consumers from the University of Michigan. The survey has a panel aspect since roughly half of the members in the sample are re-interviewed after six months. Using a probit model, the authors find that both an increase in the amount of news perceived by agents and an increase in the volatility of inflation will increase the probability that an agent changes her expectations during these six months. An agent might see the high volatility as a signal for a shock, which implies that the rewards of updating increases.

By making purchases, the agent automatically gathers information regarding the volatil-ity of prices. Using similar reasoning, it seems likely that high volatilvolatil-ity in household-specific inflation might signal a shock to an agent which increases the reward of updating his/her inflation forecasts.

Hypothesis 4 The likelihood that an agent updates his/her inflation forecasts increases if the volatility of his/her household-specific inflation increases.

2.2.3 Information channel

The different-forecasting/information approach assumes that an agent faces constraints gathering information or calculating their expectations. This paper will focus on the con-straints which are associated with gathering information. Similar to the previous approach, it is assumed by economists, that gathering information is costly. These information costs, sometimes also defined as search costs, are costs faced by individuals when acquiring infor-mation and a common way to think about this is that it takes time for an agent to search for information (Diehl et al. (2003)).

In contrast to the previous channel, this approach assumes that not all information has the same costs. Some, higher quality information, requires more time to gather (for example reading the news about monetary policy) than other types of lower quality information (for example talking with a neighbor about the economy). Given these differences in prices, agents will determine the optimal amount of information they want to purchase.

This line of reasoning is used in the expectation formation model of Malmendier and Nagel (2015), which uses both the approaches of different-forecasting and learning models. Agents will only use data about inflation from when they were alive because looking up historical data is more costly than using the data faced by individuals during their lifetime. These authors formalize this idea by building on existing adaptive learning algorithms in the macroeconomics literature in which agents estimate forecasting rules from historical data.

They find evidence, in the Consumer survey of the University of Michigan, that agents are more strongly affected by inflation realizations during their lifetime than by historical data of before their birth. Young individuals react more strongly to an inflation surprise than older individuals, who already have a longer data series accumulated in their lifetime histories. As a result, different generations disagree about the outlook for inflation.

This paper makes two assumptions in this channel. Firstly, the information cost of household-specific inflation is lower than the costs of obtaining information about the official forecasts. This is because information regarding the household-specific inflation is acquired automatically by households when purchasing goods, in contrast to gaining information about the current level of inflation and the forecasts of professional investors, as this requires the agent to take action (read a newspaper for example). And secondly, we assume that the cost of acquiring information is similar for all agents. As there is no model in the literature available which is able to do so and creating such a model is outside of the scope of this paper.

In line with D’acunto et al. (2019) we expect that if an agent purchases higher costs information, that this information will ’crowd out’ the cheaper (less sophisticated) infor-mation. We would, therefore, expect that people who indicate that they read inflation forecasts will rely less upon information acquired when doing purchases themselves and more on professional forecasts.

Hypothesis 5 Agents who read news regarding inflation will rely more on the official pub-lished inflation rates crowding out the information of the inflation rates faced when purchas-ing goods.

3

Data

3.1

Household-specific inflation

The data, regarding inflation, in this paper is drawn from the 1996-2011 Consumer Budget Surveys (CBS) of the Labor Bureau of Statistics (LBS). This survey asks approximately 500 randomly selected consumers to keep a diary of all their purchases for two weeks; after which a new sample is drawn. On average, a yearly number of 12.740 consumers held a diary for two weeks during the period covered in this paper.

The reported purchases are classified into 710 different product groups, which are divided into 60 product categories. Besides the purchasing data, the microdata of this survey also contains information related to household characteristics, such as age, income, and the number of persons living in the household.

Since the LBS only reports the average price of each different product category.4 _{It is not}

possible to use price indices specific to the sub-group of which we calculate the household-specific inflation rate. This ’common price indices assumption’ (Flower and Wales (2014)) implies that all purchases in a particular category of goods have the same price changes and that all individuals face these same price changes equally. For example, the price changes of food bought by a household with an income of above 70.000$ is the same as the price changes faced by a household with an income below 15.000$. Despite the fact that they could buy different goods within these categories, with different price changes or that they could face different prices for the same good. This assumption, which is common in

4_{A detailed overview of how these prices are derived can be found in the ’handbook’ provided by Jacobs}

comparable researches such as Kalwij et al. (2018), is made because of the lack of data which can link the expenditures documented in the CBS with the price paid for a specific good. This assumption means that the variety of experienced inflation among consumers is solely determined by the differences in consumption bundles and not by the different prices paid for a product. Kaplan and Schulhofer-Wohl (2017) can relax this assumption by using the (commercial) Nielsen scanner data, which contains both the exact good purchased by an individual and the amount paid. These authors find that relaxing this assumption increases the differences in experienced inflation among consumers in the United States but that, albeit with a larger size, the results are similar to researches which use this constraint.

3.1.1 Method

The household-specific inflation experiences of this paper are calculated using a chained-Laspeyres index; a method which is comparable to the official CPI statistic of the LBS. The benefit of using a chained-index, where the consumption bundle is based on current expenditures, compared to the regular index, where the consumption bundle is based on the expenditures in some base period, is that the chained-index allows for some degree of substitution effect between different categories (Cage et al. (2003)). In other words: individuals can reallocate their expenditures to another category if they face a change in prices.

The Price index for household i in year t is given by:

log(Pit) = J

X

j=1

w_itj log(pj_t) (1)

where wj_it is the share of total expenditures that household i spends on good j, J is the number of goods and pj_t is the price index of good j in year t. The inflation rate that a household i in year t experiences can be approximated, as for example also performed by Kalwij et al. (2018), by the following weighted inflation index :

J

X

j=1

wj_it∆ log(pj_t) (2)

where the price of good j is approximated by:

∆log(pj_t) = log(pj_t) − log(pj_t−1) (3) It uses actual consumer expenditure estimates from both the current and previous months to weight the expenditure shares indexes, as explained in the previous section, as a means of accounting for consumer substitution between item categories (Greenlees (2011)). The previous month is also included in order to decrease the effect of outliers, a strategy which is also used in the calculation of the official CPI.

w_itj = q j it+ q j it−1 PJ j=1(q j it+ q j it−1) (4)

level which aggregates all individuals. Section 3.3, in which the inflation rates will be extrapolated to the Michigan survey containing expectations, will explain in further detail how these subgroups are defined.

3.1.2 Overview of household-specific inflation

Figure 1 shows a comparison between the CPI as calculated by the LBS and as by this paper. The left panel shows the accumulated CPI from 1997 until 2011, and the right panel is the difference between the monthly inflation as calculated by the CPI and by this paper. Despite the use of a similar approach, the monthly change in CPI shows a common trend (0.98 correlated) but is not the same. The right panel shows that the difference in the calculated CPI and the official CPI is almost white noise but that the mean is lower (on average 0.004 %). This difference is mainly caused by the lower share of shelter in the share of expenditures in this paper compared to the official CPI. The price of shelter increased continuously at a high rate during this period and was relatively stable during the crisis. There are two possible explanations for the difference in shelter in the share of expenditures: not all data of the CES is made available in the public-microdata (I) and not all observations are equally weighted in the CPI of the LBS (II). The LBS does this latter as, the sample drawn in the Consumer Budget Survey, is not a perfect representation of the US population. This last issue, however, does not form a problem for this paper since the interest of this research is not to calculate the overall inflation level. By using the characteristics of a household in calculating their household-specific inflation, we automatically overcome the sampling bias of the survey.

1997 2000 2002 2005 2007 2010 100 120 140 Official CPI Calculated CPI

(a) Accumulated inflation over time

1997 2000 2002 2005 2007 2010 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 Difference in inflation Reference line

(b) Difference in monthly inflation (%)

Table 1: Household inflation experiences by household characteristic between 1997 and 2011

Household characteristic Average annual inflation Minimum Maximum Sex Male 1.86 -1.853 4.43 Female 1.85 -1.145 4.29 Age 18-29 2.08 -1.50 4.53 30-39 1.92 -1.85 4.28 40-49 1.90 -2.10 4.46 50-64 1.72 -1.50 4.31 65 and older 1.86 -0.74 4.15 Number of children 0 1.85 -1.22 4.48 1 1.91 -1.40 4.18 2 1.90 -1.95 4.44 3 or more 1.98 -2.05 4.06

Number of persons in household

Table 1 presents the calculated inflation rates for households with different character-istics. The average annual inflation rate is similar for males and females and between the different regions. However, there are substantial differences in the minimum and maximum level of inflation over time. The experienced inflation rate decreases with age but increases sharply for individuals above 65. This is mainly caused by an increase in health expendi-tures of which the price increased substantially between 1997 and 2011. We see a similar declining pattern in the inflation rates among income level, where lower income levels ex-perience a higher level of inflation. These findings are in line with previous estimations for household-specific inflation in the Netherlands (Kalwij et al. (2018)), the UK (Flower and Wales (2014)) and the United States (Kaplan and Schulhofer-Wohl (2017)).

This research does not find a relation between the number of persons in a household and the level of inflation; in contrast to these previous papers which find a negative relation.

3.2

Measuring inflation expectations

The data on inflation expectations used in this paper comes from the ’Survey of Consumers’ which is conducted by the University of Michigan since 1979. Every month a random group of approximately 400 US individuals is interviewed via telephone and asked a set of questions regarding their financial situation and their expectations of future economic con-ditions.Besides these questions, the respondents answer questions about their demographic conditions as well (e.g., age, sex, number of persons in the household). After six months, half of the sample is re-interviewed and asked the same set of questions.

The question regarding inflation expectations, which is of main interest to this paper, is: Question 1 By about what percent do you expect prices to go (up/down) on the average, during the next 12 months?

This paper looks at the period from 1996 until 2011 and will only look at people who are re-interviewed and have the same characteristics during both interviews. Both these restrictions are required for the calculation of household-specific inflation. The time-span is restricted to these years, even though a more extended time is available for this survey, since the household expenditure Survey microdata, needed for calculating inflation, is only available from 1996 until 2011.

This paper uses only re-interviewed individuals with the same characteristics during both interviews, because, as explained in the previous section, the household-specific inflation is determined based on specific characteristics of a household. Some of the properties of a consumer which determines their experienced prices can change over time (such as income, region and household size) which implies that we need to observe these characteristics during every period in which we want to calculate its experienced prices. Since this survey does not follow the same individual over time, we do not know the characteristics of an agent before his/her first interview. Therefore we cannot match a household-specific inflation with the first observation alone. For example, if we observe an income of 40.000$ for an individual in the first interview, we do not know whether his income was lower, equal or higher in the months before the interview and hence how to match it to a household-specific inflation rate. Therefore we will only use the second interview where we assume that, if an agent has the same characteristics during both interviews, these characteristics would also be the same in the months between these interviews.

of the question. Question 1 asks respondents about ”prices” and not inflation which might cause some individuals to think more about the price change in a specific group of products (for example gas prices) and other to the general inflation level (De Bruin et al. (2012)). However, despite this possible problem, the survey remains interesting from an academic and policymaker level. As Armantier et al. (2015) find in an experimental setting, that agents act on the beliefs they express when answering this question.

Table 2: The 12-month ahead inflation forecast by household characteristic between 1997 and 2011

Household characteristic N Mean Standard Deviation Sex Male 12346 2.89 3.53 Female 12958 3.46 4.06 Age 18-29 2654 3.05 3.70 30-39 4473 2.90 3.60 40-49 5906 3.15 3.75 50-64 7519 3.26 3.84 65 and older 4752 3.45 4.12 Number of children 0 16217 3.24 3.85 1 3605 3.12 3.68 2 3614 3.04 3.72 3 or more 1754 3.10 4.04

Number of persons in household

1 6661 3.29 4.01 2 15051 3.10 3.74 3 2727 3.28 3.74 4 or more 865 3.37 3.83 Income level 0 - $19.999 2828 3.81 4.55 $20.000 - $39.999 5143 3.53 4.10 $40.000 - $69.999 7127 3.13 3.69 $70.000 or more 10206 2.87 3.50 Region North 5184 3.03 3.80 East 6845 3.29 3.76 South 4681 3.08 3.88 West 8694 3.25 3.83 Total 25304 3.18 3.82

There are substantial differences in the inflation expectations among consumers. This paper will only describe these differences briefly as the heterogeneity found in the Michigan Survey is already discussed in detail by precious researches (see for example Mankiw et al. (2003)). The following trends can be found by looking at the mean of inflation expectations of various subgroups of the population. The mean of inflation expectations are substantially higher for males (2.89) compared to females (3.46), differs between regions, increase with age and decreases with income. The effect of household-size and the number of children in the household follows a trend which is less clear.

In the following section, we will elaborate on the procedure which is used to match the household-specific inflation rates to these inflation expectations.

3.3

Matching of inflation expectations and household specific

in-flation

The final calculation of the household-specific inflation and the matching to an individual of the Michigan survey of consumers is based on three steps. First, a set of household characteristics, which is observed for both the households of the Expenditure survey of the LBS and the households of the Michigan survey is determined. Secondly, based on these characteristics, a household-specific inflation is calculated as explained in section 3.1. Finally, this inflation rate is matched to an individual of the Michigan survey with the same characteristics.

Using this approach, we end up with six monthly inflation rates (which depend on the set of characteristics) for each household between their first and second interview. We only observe the inflation between these two moments since, as explained in the previous section, we only observe the time-variant characteristics for these two moments.

In the rest of this paper, the six monthly inflation rates are transformed to an annual rate, to increase the ease at which the coefficient can be interpreted during the empirical analysis. The transformation of monthly inflation to the yearly inflation is mathematically expressed as: πith = 6 Y i=t0 πi,t x 2 (5)

The six monthly inflation rates are multiplied with each other to create a semi-annual inflation rate and multiplied by two to derive an annual inflation rate.

The first step of the above process, when the set of characteristics is determined, involves a trade-off between acquiring a more ’fine-grained’ level of household-specific inflation, on the one hand, and a decrease in reliability on the other hand. By increasing the number of characteristics, the consumption bundle used to calculate the inflation becomes more ’fine-grained’ and will match more of the observable characteristics of a household of the Michigan Survey. This decreases, however, the reliability of the household-specific inflation since the number of individuals on which the household-specific inflation is based decreases as less individuals are matching all imposed characteristics. The smaller sample size decreases the reliability of extrapolating the results found using the expenditure survey of labor statis-tics to the overall population. A process performed when matching the household-specific inflation to the individuals of the Michigan Survey.

measures based on cohort and gender (IS), age-cohort and gender (AS), income-cohort and age-income-cohort (IA) and one on all of these three variables (IAS).

This paper does not use inflation measures based on more than three characteristics because by adding more subgroups the number of individuals which are available to calculate the inflation becomes too few. The household-specific inflation is no longer available for all observations of the Michigan survey as there are no individuals matching all characteristics in one of the six months for which the inflation is calculated in the case of a inflation rate based on four characteristics: during some months no households are reporting their transactions in the expenditure survey of the LBS which fulfill all criteria (I). Moreover, there is evidence which suggests that the inflation numbers are unreliable for the months in which we can calculate the inflation (II). The average standard deviation of household-specific inflation based on respectively two, three or four characteristics are 7.1, 7.7 and 11.7. The standard deviation of inflation measures based on more characteristics increases substantially, which might indicate that, because of the smaller sample size per inflation group, the inflation rates are more sensitive to outliers. Since there is no formal threshold when the inflation rate is reliable, the inflation measure which uses three characteristics is provided in this paper. However, we should be careful to draw conclusions based on this measure because, as expressed in this section, the possibility exists that it no longer is reliable to extrapolate it.

There are two reason why income-cohorts, age-cohorts and gender are used in the spec-ification based on two characteristics. First, the empirical work of Kaplan and Schulhofer-Wohl (2017) finds that income and age are the main explanatory characteristic of households in explaining variation in experienced inflation among households when the assumption of ’common price indices assumption’ is relaxed. Which indicates that our results would still be valid when this assumption is not imposed. Secondly, it follows from table 2 that the expectations of households differ most between agents of different genders, age-cohorts or income-levels. This makes it interesting to base our household-specific inflation on these measures if we want to explain the heterogeneity in expectations among consumers.

4

The effect of household specific inflation on inflation

expectations

4.1

The base line model

The baseline model studies the relation between the measures of household-level inflation and households’ expectations regarding general inflation in the subsequent 12 months. This relationship is estimated using the following linear specification by ordinary least squares:

πe_i,t= α + βπ_ith+ θπ_t−1e + Diγ + Niζ + ηt+ εi (6)

Where πe

i,t is the expected 12 months ahead inflation at time t and πhit is the explanatory

variable in which this paper is interested: the household-specific inflation experienced by a household between period t and t − 1. We include πe

i,t−1, the expected inflation at time

Furthermore, year-dummies (ηt) are added to allow the intercept to have a different

value in each year and account for effects that affect all individuals in the sample similarily but which varies per year. Year-dummies are used instead of month-dummies to enable the possibility of variables which vary only over time and not between individuals. This way, we can observe the effect of, for example, the general inflation rate on expectations. Which, when compared to the household-specific inflation, can give us an idea of the economical relevance of both measures (these results are shown in column (3)).

A vector of demographic variables including income, age, sex, household size, and edu-cation of individual i is added in all specifiedu-cations. This vector controls for time-invariant effects of demographic aspects (gender) and time-variant individual effects such as income and age. Omitting these confounding variables leads to inflated results for our dependent variable of interest (Spanos (2006)).

Two robustness tests are performed to test the sensitivity of our baseline specification. First, we will add a vector of variables Niregarding other sources of information which can

be used by agents to form their inflation expectations besides the price changes experienced when purchasing goods. The vector contains variables such as a binary dummy which is one if a household reports to have heard news regarding inflation in the past six months, aggregated (core) inflation levels as published by the Bureau of Labor Statistics and inflation forecasts by professionals. Related researches use the inflation forecasts by professionals as a proxy for gaining full information as it is the best available forecast even though it might also potentially suffer from biases (Dr¨ager and Lamla (2017)).

Secondly, we will add the aggregated inflation level (which is also depicted in Figure 1), which is very similar to the non-core inflation as calculated by the Bureau of Labor Statistics, including food and energy price changes.5 _{By controlling for the general inflation rate, we}

can test whether the household-specific inflation rate has any direct effect on expectations and that any effect is not the result of the high-level of correlation it has with general (non-core) inflation. For example, we could still find significant values results for the household-specific inflation, even though it does not have any explanatory power, if households in reality base their expectations on the general inflation level, and the household-specific inflation for various demographic groups only follows a random walk with as the mean the general inflation. The summary statistics of all these variables are available in appendix A. Table 3 reports the results for the estimation of equation 6. The first column shows that all measures of household-specific inflation are positive and significant at the 1 % level. The coefficients for the IS, IA, AS, and IAS model are respectively 0.067, 0.073, 0.035, and 0.040. The economic interpretation of the coefficients is, for example, if the annual household-specific inflation (based on IAS) increases with 1 percentage point than the twelve month ahead price expectations of a household increases with 0.04 percentage point. There are, however, differences between the magnitude of the effects found using the different measures. The three measures based on two characteristics (IA, IS, AS) imply that a large part of the effect of household-specific inflation is driven by gender. As the size of IS and AS (0.067 and 0.073), which are based partly on gender, is more than double of the IA measure, which is not based on gender (0.035). The measure which includes all three characteristics (IAS) has a coefficient of 0.040, which is slightly higher than the IA measure but substantially lower than the IS an AS measure. This contrasts our intuition that having a more fine-grained inflation rate should have more explanatory power. A possible explanation could be that the more fine-grained nature of this last inflation measure decreases the reliability of the

5_{See section 3.1.2 for a comparison between the inflation level as calculated by the Bureau of Labor}

household-specific measure, which leads to lower coefficients.

In the second column, the vector Ni is added as a robustness test. All coefficients

remain significant at the 1% level and the coefficients, even though slightly lower, remain similar. The IAS measure now has a coefficient of 0.029, and the three measures which use two characteristics IS, AS and IA have coefficients of 0.053, 0.057, and 0.025 respectively. These results are in line, albeit substantially smaller in size, with the findings of D’acunto et al. (2019). These authors find a coefficient of 0.17 for the effect of experienced inflation on the expectations of consumers. The difference in size is probably due to the fact that these authors do not have to impose the common price indices assumption as they use a different data-set, as explained in section 2. This means that the variety in household-specific inflation is higher. Also, because they can link the expenditure data directly to the price expectations of an individual, they do not have to extrapolate the inflation rate. This means that, in contrast to the current research, they do not have to match an inflation rate based on observable characteristics but rather have an inflation rate, which is directly linked to a consumers’ inflation expectations.

The results of the third specification, which includes the aggregated inflation level (in-cluding food and energy prices), shows that the coefficients and the significance level changes and that these effects differ between the household-specific inflation measures. The IAS measure is no longer significant; the measure for general inflation in this specification is significant at the 1% level and has a value of 0.095. The household specific inflation and general inflation for the IS and IA measure are both significant at the 1% level and have coefficients of 0.036 and -0.013 respectively. The AS measure is significant at the 10% level and has a coefficient of 0.022. The general inflation is significant in all three specifications at the 1% level but the size varies substantially between the different measures. On average, however, the coefficient of general inflation is larger than the coefficient of household-specific inflation.

We should, however, be careful in interpreting the coefficients of the four household-specific inflation measures and the general inflation level. The high correlation (0.76-0.90) between household-specific and aggregated inflation can lead to multicollinearity. The vari-ance inflation factor (VIF), which is a formal test for multicollinearity, has a value for the IAS, IS, AS and IA measure of respectively: 2.5, 6.15, 9.68 and 5.28. It follows from our specification of the household specific-inflation that the measures which consist of more characteristics have lower VIF values as we allow the measure to vary more between indi-viduals. This way, the measure can deviate more of the (aggregated) general inflation level (which is the same for all households) and therefore the VIF will decrease.

The econometric literature describes several threshold values for the VIF (O’brien (2007)), but generally, a value 5 to 10 is considered to be a reason to treat the coefficient with cau-tion as this indicates, to some degree, the presence of multicollinearity which decreases the reliability of the estimates. This means that it is not acceptable to draw firm conclusions regarding the size of the coefficients of the IS, AS or IA inflation measures (as these VIF values are above 5).

As a further robustness test of the above results, we also controlled for the effect of aggregated inflation by adding time-dummies for all periods. These dummies allow the intercept to have a different value in each period and, therefore, controls for all effects which are constant across individuals but which changes over time (such as the aggregated inflation level). The results of this robustness test are in appendix B. The VIF values increase further (all above 10), which means that the coefficients are severely affected by multicollinearity. Therefore, we should be cautious when interpreting the results. The sign and significance-level of the inflation measurements in this regression are similar to the regression including general inflation. This is a slight reconfirmation of the conclusion which was drawn in the previous paragraph: household specific inflation has a significant impact when we control for the general inflation level.

4.2

Heuristics

4.2.1 Frecuency-based consumption bundle

In section 2.2.1, this paper formed the hypothesis that ”frequency-baskets” are more im-portant than ”purchase-baskets” when specifying the consumption bundle used to calculate the inflation-level. To test this hypothesis, we recalculate the household-specific inflation by transforming equation 4 to:

wj_it= f j it+ f j it−1 PJ j=1(f j it+ f j it−1) (7)

Where f_itj is the frequency of how often good j is bought by household i in period t. The equation of the baseline is re-estimated (equation 6), including all control variables, for both the household-specific inflation based on the frequency basket and the purchase share basket.

Table 4 shows the results of the OLS regression of equation 6 with the frequency based basket. In this regression, a frequency measure of general inflation is not included in contrast to the final specification of the previous regression. When using both the household-specific inflation and a general inflation frequency based-basket, the former becomes insignificant or negative for all specifications. The VIF value, however, is between 11 and 40, for both the aggregated and household specific inflation, which indicates that there is severe multi-collinearity making the results unreliable. Dropping the general inflation level solves this issue but implies that we cannot rule out the possibility that the general inflation level (partly) drives these results.

The effect of the frequency-basket is positive and significant at the 1% significance level for all four specifications of the household-specific inflation. The coefficient of the frequency-basket inflation is substantially higher than the coefficient of the expenditure-frequency-basket, indi-cating that the inflation expectations of individuals react more strongly to changes in the frequency-basket inflation. This finding is in line with our second hypothesis.

When including both the frequency-basket and the expenditure-basket, both inflation measures decrease in size but remain significant at the 1% level, implying that individuals either: use both measures of inflation when making their predictions (I) or that it differs between agents which model they use (II). However, more research is needed to find which of these effects is leading.

4.2.2 Extreme changes in household-specific inflation

Table 5: The effect of easier to recall changes on inflation expectations

max inf. gasoline

househould inflation IS AS IA IAS IS AS IA IAS Household Inflation 0.051*** (0.005) 0.056*** (0.006) 0.022*** (0.004) 0.028*** (0.004) 0.034*** (0.006) 0.033*** (0.006) 0.0068** (0.004) 0.013*** (0.004) maximum infl. 0.17 (0.07) 0.14** (0.07) 0.16*** (0.05) 0.07 (0.056) gasoline 0.016*** (0.003) 0.017*** (0.003) 0.024*** (0.003) 0.023*** (0.003) Information X X X X X X X X Demographics X X X X X X X X Year FE X X X X X X X X Nobs 25068 25068 25068 25068 25068 25068 25068 25068 R2 0.10 0.10 0.10 0.10 0.11 0.11 0.11 0.11

Standard errors in parentheses p<0.10, ** p<0.05, *** p<0.01

IS = income-cohort,sex; AS = age-cohort,sex; IA = income-cohort,age-cohort; IAS = income-cohort,age-cohort,sex

The second hypothesis regarding heuristics predicts households to overestimate periods with more extreme price changes as households find it easier to recall these levels of inflation when making forecasts. We account for this effect by including the (monthly) household-specific inflation which deviated most from the mean value of the six monthly inflation rates experienced by the household prior to making his/her forecast. Or in mathematical notation:

πmax_it = |¯π_ith− πi,t| (8)

Furthermore, the price change in gasoline is added. Following our third hypothesis, and explained by Binder (2018) and Bruine de Bruin et al. (2011), we expect households to overweigh the experienced price changes of gas when making inflation forecasts. As these prices are volatile, frequently purchased, and visible. Again, these two hypotheses are tested using equation 6. Table 5 shows the result of both effects. The most extreme value of household-specific inflation is positive and significant at the 1% significance level and is consistent for all measures based on two-characteristics (IA, IS, and AS). A one-percentage-point increase of the most extreme monthly inflation rate (for the IA measure) in the past six month increases the 12-months ahead expected price change with 0.14. Or, if recalculated to a yearly inflation rate: 0,11.

The ’extreme’ value of the IAS measure is not significant, which, as explained in the previous section could be caused by the unreliability of this measure.

driver of the effect of household-specific inflation on inflation expectations. The size of the gasoline measure compared to the effect of household-specific inflations is sensitive to the different inflation measures. In the case of the IS and AS measure the effect of gas prices (0.016) is half of the effect of household-specific inflation (0.034). In the case of the IA and IAS measure, the effect of gasoline price changes on inflation is comparable or smaller than the effect of household-specific inflation.

4.3

The updating channel

Table 6: The effect of household specific inflation on the likelihood of updating expectations

Household inflation IS AS IA IAS Variance Household inflation 0.011

(0.029) -0.021 (0.044) 0.018 (0.027) -0.020 (0.025) Variance Professional inflation 0.028***

(0.007) 0.033*** (0.009) 0.027*** (0.007) 0.032*** (0.007) News change 0.051*** (0.018) 0.051*** (0.018) 0.051*** (0.018) 0.51*** (0.19) Demographics X X X X Nobs 25286 25286 25286 25286

Standard errors in parentheses p<0.10, ** p<0.05, *** p<0.01

IS = income-cohort,sex; AS = age-cohort,sex; IA = income-cohort,age-cohort; IAS = income-cohort,age-cohort,sex

The empirical validation of the updating channel is done using the same method as proposed by Dr¨ager and Lamla (2017). In their empirical research, which uses the same dataset as this paper, the authors try to endogenize the decision of agents when to update their expectations. For this, they use a probit model where the explanatory variable is a binary dummy which is one if the expected inflation reported by an individual differs between the two interviews of the Michigan Survey or zero if a respondent gives the same price expectation in both interviews. This dummy gives the lower-bound of measuring if an agent has updated his/her expectations in the six months between both interviews as it does not take the possibility into account that an individual updated his/her expectations but ended up with similar inflation forecasts. In such a case, even though he/she updated her expectations, the dummy will be zero.

Furthermore, the same control variables are included as in the specification of Dr¨ager and Lamla (2017). A dummy which is one if a person reported in the first interview that they didn’t hear news regarding inflation and during the second interview that they did, a measure for the volatility in CPI as reported by the labor bureau of statistics and a variable for the volatility in the inflation forecasts of professional forecasters. The theoretical relevance of these control variables is explained in section 2.2.2: an increase in the volatility might signal a shock, which increases the reward of updating as the time exposed to this shock decreases. The results are summarized in table 6.

our hypothesis that an increase in the volatility of household-specific inflation increases the probability of an agent updating his/her expectations. Moreover, the result indicates that we do not find evidence that household-specific inflation plays a role in models using information-rigidities.

4.4

Information channel

Table 7: The interaction effects between hearing news and household specific inflation

Interaction effects of news household inflation IS AS IA IAS Household Inflation (HI) 0.052*** (0.005) 0.056*** (0.006) 0.024*** (0.004) 0.029*** (0.004) Professional inf. (PI) 0.39*** (0.058) 0.42*** (0.059) 0.53*** (0.057) 0.52*** (0.058) Increase News (Inews) 0.98*** (0.15) 0.91*** (0.15) 0.91*** (0.13) 0.98*** (0.15) Decrease News (Dnews) -0.78** (0.027) -0.74** (0.027) -0.78** (0.027) -0.78** (0.027) HI x Inews -0.015 (0.013) -0.016 (0.013) -0.010 (0.010) -0.016 (0.012) HI x Dnews -0.019 (0.029) -0.017 (0.029) -0.016 (0.028) -0.016 (0.028) Information X X X X Demographics X X X X Year FE X X X X Nobs 25068 25068 25068 25068 R2 0.10 0.10 0.10 0.10

Standard errors in parentheses p<0.10, ** p<0.05, *** p<0.01

IS = income-cohort,sex; AS = age-cohort,sex; IA = income-cohort,age-cohort; IAS = income-cohort,age-cohort,sex

For this channel, we look at how the ’purchase’ of more information changes the use of household-specific inflation by an individual. To do so, we use a binary dummy which is one if a consumer indicates to have heard that inflation will increase and a similar dummy which is one if a person reports having heard news that the prices will decrease. These dummies are interacted with household-specific inflation to see whether individuals who have heard news regarding inflation will rely less on the household-specific inflation. According to our hypothesis, we expect that individuals will rely more strongly different sources of information when they have heard news since this information is of higher quality, and that the use of household-specific inflation will then decrease.

has heard news regarding an increase of inflation in the past six months, he/she will increase their inflation expectations by 0.985 The coefficients of these effects are much higher than the coefficients of household-specific inflation. This importance of news is in line with the work of Pfajfar and Santoro (2013) who find that the individuals update their expectations if they hear news regarding inflation.

5

Conclusion

This study analyzed the effect of price changes experienced by households when making pur-chases (household-specific inflation) on the formation of inflation expectations of consumers. Until recently, most economists overlooked the impact of household-specific inflation on con-sumers’ expectations because they assumed the variety in experienced inflation rates across consumers to be limited. This research, however, finds evidence which indicates that we should reject this assumption and that household-specific inflation plays a substantial role in the formation of short-term price forecasts of consumers. This effect, as the household-specific inflation differs among agents, can explain part of the heterogeneity of consumers’ price forecasts.

In the theoretical analysis, three channels were identified, based on previous literature on expectation formation, how household-specific inflation plays a role in the price forecasts of consumers. Household-specific inflation might: impact the heuristics used by agents when calculating their inflation forecasts (heuristics), influence the probability that individuals update their inflation expectations (updating), and might be stronger for individuals who do not read about official inflation forecasts (information). Each of these channels finds its origin in a particular rejection of FIRE.

These channels are empirically validated on a sample of US consumers from 1997 to 2011. Household-specific inflation measures were calculated using micro-expenditure data of the Labor Bureau of Statistics and matched to the Michigan Survey of consumers, which contains a question regarding expected 12 month-ahead price changes, based on observ-able properties of each agent. This way, inflation indices were calculated based on three characteristics: age, gender, and income.

The empirical results find evidence that the effect of household-specific inflation is sig-nificant and positive and that these results are robust when controlled for a large number of demographic factors (income, age, gender, education, and family size) and other sources of information (inflation forecasts of professionals and core inflation). When controlled for the non-core inflation level, mixed effects of the household-specific inflation on expectations are found. Furthermore, because a moderate level of multicollinearity is detected between the general non-core inflation level and household-specific measures, we cannot draw firm conclusions regarding these coefficients. Implying that we do find evidence in favor of, but cannot with certainty, reject the common assumption that there is no difference between the effect of general inflation level (including electricity and food prices) and the household-specific inflation level on inflation expectations of consumers. Further research is needed to increase the certainty at which we can reject this assumption.

that (the volatility in) household-specific inflation increases the probability that an agent updates his/her expectations in a information rigidities model (updating-channel) and we do not find evidence that perceiving news changes the role of household-specific inflation (information-channel).

There are two main limitations to this research, which both stem from the availability of data. First, we have to make the ’common price indices assumption’; the assumption that the change in the price of a good is the same for all goods in the same product cat-egory and that the prices faced for a particular product are similar for all agents. This restricts our measure of household-specific inflation severely as the only source of variation between agents are the differences in consumption bundles. Secondly, because the expendi-ture and expectation data come from two different data-sets, we are forced to extrapolate the household-specific inflation and match these inflation rates based on observable character-istics of the consumers in the Michigan Survey of Consumers. This has two disadvantages. First, differences in the household-specific inflation which stem from unobserved households characteristics are not taken into account, which means that our household-specific inflation only accounts for a limited part for all price changes experienced by households when making purchases. Moreover, the extrapolation of household-specific inflation becomes unreliable when we create too fine-grained inflation indices (based on more than two characteristics) as the number of observations on which the inflation is based becomes too low. This means that we can only test our hypotheses based on household-specific inflation rates, which are strongly restricted in the number of observable characteristics they take into account. This results in a high correlation between the household-specific inflation and general inflation, limiting our ability to draw strong conclussions when both these measures are included in the regression.

Despite these limitations, this research is still valuable for policymakers. This paper finds that central banks should not only look at the effect on aggregated inflation level (the policy-target) when using policy instruments but should also look at the effect of household-specific inflation. Differences in experienced price changes, when making purchases, affect the inflation forecasts of individuals, which has an impact on the policy target. Incorporating this effect into models might, therefore, increase the accuracy at which central bankers can predict the outcome of their policy.

This research suggests that more research is needed to explain ’how’ household-specific inflation affects price forecasts’ of consumers as only little evidence is found in this paper how this effect fits into current expectation formation models. I suggest two possible ways of doing so. Firstly, economists can look at the impact of household-specific inflation on expectations in adaptive learning models; a class of models which is not investigated in this paper because of the lack of data. And secondly, researchers could extend the work of this paper on the information channel. This paper assumes that the ’cost’ of purchasing information is the same for all individuals, an assumption which is rejected by previous ex-perimental research. Future research could relax this assumption to see whether individuals who face higher costs of buying more sophisticated information tend to rely more heavily on household-specific inflation.

References

their beliefs?” International Economic Review, 56, 505–536.

Armantier, Olivier, Scott Nelson, Giorgio Topa, Wilbert Van der Klaauw, and Basit Za-far (2016), “The price is right: Updating inflation expectations in a randomized price information experiment.” Review of Economics and Statistics, 98, 503–523.

Badarinza, Cristian and Marco Buchmann (2011), “Macroeconomic vulnerability and dis-agreement in expectations.”

Bernanke, Ben (2007), “Inflation expectations and inflation forecasting.” Technical report, Board of Governors of the Federal Reserve System (US).

Bernanke, Ben, Michael T Kiley, and John M Roberts (2019), “Monetary policy strategies for a low-rate environment.”

Binder, Carola Conces (2018), “Inflation expectations and the price at the pump.” Journal of Macroeconomics, 58, 1–18.

Branch, William A and Bruce McGough (2018), “Heterogeneous expectations and micro-foundations in macroeconomics.” Handbook of Computational Economics, 4, 3–62. Brown, Roger and James Kulik (1977), “Flashbulb memories.” Cognition, 5, 73–99. Bruine de Bruin, W¨andi, Wilbert van der Klaauw, and Giorgio Topa (2011), “Expectations

of inflation: The biasing effect of thoughts about specific prices.” FRB of New York Staff Report.

Bruine de Bruin, W¨andi, Wilbert Vanderklaauw, Julie S Downs, Baruch Fischhoff, Giorgio Topa, and Olivier Armantier (2010), “Expectations of inflation: The role of demographic variables, expectation formation, and financial literacy.” Journal of Consumer Affairs, 44, 381–402.

Cage, Robert, John Greenlees, and Patrick Jackman (2003), “Introducing the chained con-sumer price index.” In International Working Group on Price Indices (Ottawa Group): Proceedings of the Seventh Meeting, 213–246, Paris: INSEE.

Calvo, Guillermo A (1983), “Staggered prices in a utility-maximizing framework.” Journal of monetary Economics, 12, 383–398.

Coibion, Olivier, Yuriy Gorodnichenko, and Rupal Kamdar (2018), “The formation of expec-tations, inflation, and the phillips curve.” Journal of Economic Literature, 56, 1447–91. Colander, David, Michael Goldberg, Armin Haas, Katarina Juselius, Alan Kirman, Thomas

Lux, and Brigitte Sloth (2009), “The financial crisis and the systemic failure of the eco-nomics profession.” Critical Review, 21, 249–267.

Curtin, Richard (1996), “Procedure to estimate price expectations.” Manuscript, University of Michigan Survey Research Center.

De Bruin, W¨andi Bruine, Wilbert Van der Klaauw, Giorgio Topa, Julie S Downs, Baruch Fischhoff, and Olivier Armantier (2012), “The effect of question wording on consumers’ reported inflation expectations.” Journal of Economic Psychology, 33, 749–757.

Demery, David and Nigel W Duck (2007), “The theory of rational expectations and the interpretation of macroeconomic data.” Journal of Macroeconomics, 29, 1–18.

Di Bartolomeo, Giovanni, Marco Di Pietro, and Bianca Giannini (2016), “Optimal monetary policy in a new keynesian model with heterogeneous expectations.” Journal of Economic Dynamics and Control, 73, 373–387.

Diehl, Kristin, Laura J Kornish, and John G Lynch Jr (2003), “Smart agents: When lower search costs for quality information increase price sensitivity.” Journal of Consumer Re-search, 30, 56–71.

Dr¨ager, Lena (2015), “Inflation perceptions and expectations in sweden–are media reports the missing link?” Oxford Bulletin of Economics and Statistics, 77, 681–700.

Dr¨ager, Lena and Michael J Lamla (2012), “Updating inflation expectations: Evidence from micro-data.” Economics Letters, 117, 807–810.

Dr¨ager, Lena and Michael J Lamla (2017), “Imperfect information and consumer inflation expectations: Evidence from microdata.” Oxford Bulletin of Economics and Statistics, 79, 933–968.

Duca, Ioana, Geoff Kenny, and Andreas Reuter (2018), “Inflation expectations, consumption and the lower bound: micro evidence from a large euro area survey.”

Flower, Tanya and Philip Wales (2014), “Variation in the inflation experience of uk house-holds: 2003-2014.” Office for National Statistics, 15.

Georganas, Sotiris, Paul J Healy, and Nan Li (2014), “Frequency bias in consumers percep-tions of inflation: An experimental study.” European Economic Review, 67, 144–158. Greenlees, John S (2011), “Improving the preliminary values of the chained cpi-u.” Journal

of Economic and Social Measurement, 36, 1–18.

Hartley, James E (2002), The representative agent in macroeconomics. Routledge.

Hobijn, Bart, Kristin Mayer, Carter Stennis, and Giorgio Topa (2009), “Household inflation experiences in the us: a comprehensive approach.” Federal Reserve Bank of San Francisco. Hommes, Cars (2011), “The heterogeneous expectations hypothesis: Some evidence from

the lab.” Journal of Economic dynamics and control, 35, 1–24.

Jacobs, Eva E (2008), Handbook of US Labor Statistics 2008: Employment, Earning, Prices, Productivity, and Other Labor Data. Bernan Press.

Kalwij, Adriaan, Robertus Alessie, Jonathan Gardner, and Ashik Anwar Ali (2018), “Infla-tion experiences of retirees.” Journal of Pension Economics & Finance, 17, 85–109. Kaplan, Greg and Sam Schulhofer-Wohl (2017), “Inflation at the household level.” Journal