1The distinction between mainstream and heterodox economics concerns, among several theoretical disputes, the way of dealing with the issue of consumption. This field still poses open questions (Drakopoulos, 2021a). Economic theories can be divided between those emphasizing the absolute aspect of consumption and those emphasizing the relative aspects of consumption, as discussed in Drakopoulos (2021b). The former group of theories focuses mainly on the optimization of consumption plans and on the intertemporal aspects of preferences; the latter group focuses on interdependent aspects of consumption. In the first case, the nature of consumption is mainly attributed to biological needs of man, or at least, in our post-industrial societies, to naturalized and rationalized needs; in the second case, consumption depends on to social processes. In both cases, there is no lack of insights from other disciplines such as psychology and neuroscience. The level of interdisciplinarity required by the postulate of interdependent preferences is higher, since it includes all those disciplines that study relationships across people, such as sociology, and not only the facets of individual choices, such as heuristics and biases. The two group of theories on consumption highlight respectively the rational character of human choices and the social one, but the two points of view are not mutually exclusive: there are articles (a milestone is Leibenstein, 1950) and theories where the homo œconomicus has a utility function that depends both on an objectively measurable quantity of money, on prices, and on a status component (conformism, social distinction, social environment/network externalities, positional goods, and so on). The literature still discusses whether people care more about the relative aspect of consumption (and income) or about the absolute one: “On one extreme, standard economic theory typically assumes, based on no empirical evidence, that only absolute income and consumption matter. On the other extreme, Easterlin (1974; 1995), Oswald (1998), and others conclude that only relative income seems to matter” (Alpizar et al., 2005, 406).

3Building upon the premises outlined above, this article aims at meticulously exploring whether there are discernible differences in the approach to consumption between articles published by a journal categorized as “heterodox” (Journal of Economic Issues), and those published by journals deemed “orthodox” (namely, Econometrica and the American Economic Review; AER hereafter). This distinction between heterodox and orthodox journals is grounded in the literature (Lee et al., 2010). Consequently, this article represents an empirical endeavor to assess whether theoretical divergences in consumption models are echoed in the linguistic patterns employed in these two publication outlets, given that language is assumed to be indicative of contents. We acknowledge the partially circular nature of its logic, thereby highlighting the importance of applying rigorous and transparent methodologies in the compilation of the dataset (corpora), as further elaborated in Section 2.

6This article is meant to be exploratory. Text mining is increasingly gaining a foothold in economic analysis and in history of economic thought, but there is little consensus on how best to deploy it (see e.g., Ash and Hansen, 2023). This article contributes therefore to the broader, emergent body of literature that applies quantitative methods to the history of economic thought (see e.g., Goutsmedt et al., 2023). The exercise performed here is a way of understanding what computational linguistics (a specific branch of textual analysis) can add to the history of economic thought, as opposed to traditional text exegesis, as well as to other purely quantitative techniques, such as the ‘bags of words’ approach (see e.g., Gentzkow et al., 2019). The latter transforms words into numbers, without the possibility of controlling for their meaning in the context. In this article, we would rather use a software, Sketch Engine (developed for linguists), that always allow to control for the exact use of the single word (or n-gram) within the whole sentence (the context), for each specific occurrence. Henceforth, our research offers an original contribution on two fronts. Methodologically, as the computational linguistics techniques employed diverge from the traditional ‘bags of words’ approach, by tracking the semantic content of texts. More substantively, the originality of this article is to explore the theory of consumption, an issue still unexplored with textual analysis, besides Baumgartner (2010)’s analysis on a corpus of marketing articles. Yet, the four results of this article are restricted to the selected journals: they cannot be generalized, although the methodology we describe here would be applicable in several research contexts.

7The article is structured as follows: the next section further clarifies the distinction between absolute and relative aspects of consumption, which constitute the main interpretative framework for the empirical investigation conducted in Section 3. Section 2 presents the methodology of this article. Section 3 discusses the four results of our synchronic analysis.

1. Delving into Absolute vs. Relative Consumption

8This section aims at convincing the reader of the relevance of the categorical variable “absolute consumption” or “relative consumption” in classifying consumption theories, whilst highlighting some tautological aspects of orthodox (“absolute”) consumption theory. In other words, this section provides historical depth to the empirical analysis conducted in the rest of the article, underlining the main historical cornerstones on which modern theories of consumption (investigated in Section 3) are built. Both orthodox and heterodox theories have indeed been enriched in many ways over time; yet, as we will try to prove in our analysis, the fundamental difference, between absolute and relative consumption, remains.

1.1 The Relative Hypothesis

10In John Maynard Keynes’ General Theory, we find the “absolute” income hypothesis, according to which consumption depends on the absolute level of current income and the marginal propensity to consume (Keynes, [1936] 2018). Despite the appearances (i.e., the word “absolute”), Keynes was convinced of the relative (social) nature of consumption, and, indeed, he linked the marginal propensity to consume to the psychological characteristics of the community. He also emphasizes the comparison taking place at wage level: workers resist a cut in monetary wages to maintain their relative position in the wage structure, and not so much to avoid a cut in their absolute income. In his words: “The effect of the combination by a group of workers is to protect their relative real wages” (Keynes, [1936] 2018, 15).

14Finally, another important author emphasizing the relative aspect of preferences and consumption is Richard Easterlin. His hedonic-satisfaction-positional treadmill concepts are of particular relevance to our classification. Indeed, they emphasize the relativity of income, consumption, welfare (or utility) and their interrelationships in different ways (Easterlin, 1974; 1995; Easterlin et al., 2010).

1.2 The Absolute Hypothesis

15Let’s turn to authors with a view of preferences as exogenous and of consumption as absolute. As mentioned earlier, Fisher (1930) dealt with the intertemporal choices between present and future consumption, providing the theoretical basis for the next trend of studies on the topic.

17Refining the intertemporal aspect of consumption by working on expectations has been a central issue in the following years. The key point of the Friedman’s permanent income hypothesis is that consumption plans do not depend on transitory income components, which is a tautology, since the transitional component is considered unpredictable (and with expected value of 0, since “good” years would compensate for “bad” years), and an unpredictable thing cannot be part of a planning by definition. In his model, agents have adaptive expectations: they need more than one-period increase in income in order to expect an increase in permanent income.

19Hall’s idea is indeed at the heart of contemporary dynamic stochastic general equilibrium (DSGE) models and of modern macroeconomic theory on consumption, since it links the present with the future and allows for recursive solutions. If the planned consumption is the same as today, then the budget constraint can be used to write current consumption in terms of wealth and present value of current and future expected income. This is the basis of the so-called rational expectations permanent income consumption function.

21Several modern theories on consumption (Bagliano and Bertola, 2004) are based on two main assumptions: (1) identical economic agents maximize an intertemporal utility function, defined on consumption levels in each period of the optimization horizon, subject to resource constraints; (2) under the conditions of uncertainty, maximization shall be based on expectations of future-relevant variables (such as income and interest rates) that are rationally inferred by agents who optimally use all the information. As mentioned above, modern macroeconomics incorporated the notion of habits in the consumption theory (Bagliano and Bertola, 2004). Habit effect on the utility function means that a high level of consumption in period t decreases utility in period t + 1 (but increases period t + 1 marginal utility). This is a way of translating the idea that, once one consumes more, he or she gets used to that level of consumption, which no longer gives the same satisfaction as in the previous period. At the same time, the utility of consuming an extra unit at t + 1 is greater if, at the time t, an increase in consumption with respect to t – 1 has already taken place. This is indeed an extension of the previous theory that now is able to fit a wider range of possible behaviors; however, this extension is still referred to an autonomous aspect of consumption and to an isolated individual. Other variables captured by modern macroeconomic models are, for instance, bounded rationality, myopic agents, precautionary savings, borrowing constraints, inconsistent and hyperbolic discount rate (for a review, see e.g., Jappelli and Pistaferri, 2017). That is, there has been an enrichment of the intertemporal theory on absolute consumption with these behavioral elements, but still avoiding preference interdependence.

22The distinction between relative and absolute consumption outlined in this section leads to the discussion of whether the individual aspects of consumption or the social ones (consumption as a socially driven process) are more relevant. In other words: Do agents choose according to a principle of universal economic rationality, or do they give a culturally mediated meaning to their consumption acts (Todorova, 2014)? This is the kind of question the recent debates on consumption still leave unanswered, since the answer differs depending on the two lineages outlined in this section. Yet, the explanation for consumption patterns by the absolute hypothesis is embodied, today, in variety of models, including those in use at independent policy institutions, like the European Central Bank (like DSGE models), or those taught in most bachelor and master courses in universities. The apparent popularity of the absolute hypothesis provides us with a further reason to delve deeper into the topic, through our linguistic analysis.

2. Methodology

2.1 Data Collection

24The journals’ choice retains a degree of irreducible arbitrariness. The first constraint to our choice was to obtain two similar-size corpora when searching for keywords in the search engines of journals. The main reason for choosing Econometrica and the AER is their good position within Lee et al. (2010)’s ranking. They are journals with a high “h-index”, and a high number of weighed average citations of articles of the last three years. Moreover, they are eminent journals in the economics field. They are also slightly different among them: Econometrica got a reputation more with methodological articles. The three journals have a general scope, that is, they are not journals focused on the issue of consumption: publications cover various topics (including empirical, experimental, theoretical and purely methodological articles), and follow approaches generally considered as different (from behavioral to evolutionary economics, up to the standard General Equilibrium framework).

2.2 General Method of Analysis

28The number of articles in the dataset makes it sensible to use textual analysis, given the difficulty one would have in reading and analyzing more than a hundred articles. A natural way of proceeding, given our research question, is to look at the high-frequency single words or/and n-grams, that is, “keywords”, then to understand how they characterize the focus corpus, and then move to study the collocations of such keywords with the use of context analysis (or “word sketch” difference).

29The “reference corpora” provides the comparative corpora used for extracting keywords. Reference corpora work thanks to their linguistic representativeness, and they serve as a comparison to the focus corpus (our articles from the AER/Econometrica and the Journal of Economic Issues) to understand which words characterize the latter. Any corpus with a miscellaneous language, and whose variety is uniformly distributed within the corpus (or distributed according to the relative frequency of use in a specific current written language) can act as a reference corpus. In our case, such reference corpora are pre-loaded on Sketch Engine (the software used in this analysis) and they have been constructed according to the strictest criteria of representativeness and balance. I have changed the reference corpus several times, to check for the robustness of the results: the basic corpus used is the English Web 2020, but I also used as reference the Cambridge Academic English corpus (specifically focused on academic language).

2.2 Metrics

31There are several ways of applying text mining to corpora—for a brief summary see e.g., Gentzkow et al. (2019) and Gupta et al. (2009). An example of textual analysis application with aims akin to this article (but with different methods) is Ash et al. (2021). Ash and coauthors provide an unsupervised method to quantify latent narrative structures in text documents, and they apply it to the United States Congressional Records, in order to analyze political and economic narratives in recent decades.

34Equations 1, 2, and 3 describe the main metrics used in the whole empirical analysis:

35\[\tag{1}TTR = \frac{number of unique tokens}{total number of tokens}\]

36\[\tag{2}KeynessScore = \frac{f_{pm_{rm_{focus}}} + n}{f_{pm_{rm_{ref}}} + n}\]

37\[\tag{3}LogDice = 14 + log_{2} \frac{2*\|w_{1} . R . w_{2} \|}{\|w_{1} . R . *\| + \| * . *, w_{2} \|}\]

38The first indicator is described here in its basic version, although the actual analysis uses its standardized version, in order to prevent the result from being biased by the corpus’ size. The ratio between the number of unique items (each word form is counted once) and the total number of tokens is assumed to measure the lexical richness (or variety) of a text. In this case, in order to avoid the problem of different number of mathematical symbols in the two corpora (articles published in the AER/Econometrica have more mathematical symbols than those from the Journal of Economic Issues), the measure of TTR is provided with lemmas (instead of words); in this way, numerical/symbolic token are excluded from the metric.

39The second scoring function is a ratio between relative frequencies: that of a target-word in the focus corpus, with respect to that for the same target-word in the reference corpus. The Keyness score is therefore a measure of specificity of some words (or n-grams) in our focus corpus. Finally, Equation 3 describes the LogDice score. Here, the intercept (the number 14 in front of the logarithm) is a way of translating up the function.

40This article is not intended to develop a predictive model, or to use a predictive model in a text mining exercise. I have just built the corpora following the representativeness criterion, broke the text down, prepared it for the software (also adding metadata), and analyzed it statistically through the scoring functions.

3. Synchronic Analysis

3.1 Lexical Richness and Technicalities

41The first results come from the descriptive statistics of the corpora and from lemmas’ relative frequencies. The AER/Econometrica corpus has 635,410 tokens and 447,462 words. The number of unique words including non-words (occurrences) is 25,199 and the lemmas are 15,611. For the Journal of Economic Issues corpus, these numbers are respectively 498,481, 391,424, 29,913 and 18,751. Some results can be directly inferred from these numbers: namely the fact that in the AER/Econometrica corpus there are more non-words than in the Journal of Economic Issues corpus. This is due to the higher presence of mathematical expressions and of Greek letters in this corpus.

42As described in the previous section, the Types Tokens Ratio (TTR) is the ratio between the unique words of a text (each word form is counted once) and the total number of tokens. TTR measures, thereof, the lexical variety or richness of a corpus—the closer to 1, the higher the lexical richness of the corpus. For the AER/Econometrica articles about consumption, this index, in its standardized version (Equation 1 above) is 0,039 (3,9%); whilst, for the Journal of Economic Issues corpus, the TTR is 0,06 (6%). Since we are dealing here with economics academic articles, it is possible that finding a high number of unique words results from the use of various mathematical symbols; therefore, a more credible measure is TTR calculated with lemmas. In this case, the TTR is 0,024 (2,4%) for the AER/Econometrica corpus, and 0,037 (3,7%) for the Journal of Economic Issues corpus. In conclusion, articles about consumption published in the Journal of Economic Issues are lexically richer than those published in the AER/Econometrica.

43Many keywords of the AER/Econometrica corpus are technical. This is not surprising, since we apply here text mining to academic articles; it is thus plausible that these texts are characterized by a significant use of specific terms from the field of economics, such as “lemma”, “equilibrium”, “convex”, “theorem”, “concave”, “elasticity”, “differentiable”, and so on. Table 1 reports some technical words, chosen among the first 100 keywords in the orthodox corpus.

Table 1. Technical Keywords in AER/Econometrica

		Frequency per million		Document frequency (DOCF)		Relative DOCF		ARF
	Word	Focus	Reference	Focus	Reference	Focus	Reference	Focus	Reference	Score
1	quantile	509.91	0.19	4	3,888	9.76%	<0.01%	30.18	3,043.78	428.8
2	lemma	569.71	0.63	20	12,102	48.78%	0.01%	84.13	9,180.25	349.6
3	theorem	826.24	4.39	26	77,230	63.41%	0.09%	161.86	59,175.53	153.5
4	multiperiod	107.02	<0.01	8	254	19.51%	<0.01%	12.13	176.53	107.2
5	concave	275.41	1.62	28	42,884	68.29%	0.05%	59.91	31,451.59	105.3
6	differentiable	110.17	0.18	20	4,710	48.78%	<0.01%	27.35	3,168.89	93.8
7	equilibrium	624.79	5.73	23	135,994	56.10%	0.17%	113.85	104,505.42	93.0
8	martingale	111.74	0.26	6	5,822	14.63%	<0.01%	12.11	4,249.75	89.8
9	covariance	149.51	0.69	12	13,972	29.27%	0.02%	23.72	10,930.50	89.1
10	recursive	180.99	1.34	16	33,257	39.02%	0.04%	31.73	25,074.49	77.7
11	monotonicity	78.69	0.06	10	1,696	24.39%	<0.01%	13.88	1,227.25	75.0
12	idiosyncratic	151.08	1.17	10	47,391	24.39%	0.06%	21.57	31,011.78	70.2
13	corollary	124.33	0.98	15	37,542	36.59%	0.05%	22.61	25,127.09	63.2
14	maximization	85.56	0.44	19	14,799	46.34%	0.02%	25.37	10,320.80	60.7
15	separable	89.71	0.50	16	15,035	39.02%	0.02%	22.32	10,719.49	60.5

Note that only technical keywords are reported in this table. Keywords have been sorted according to their Keyness score.

44The first column refers to the relative frequencies in the focus and in the reference corpus respectively. The second column reports the share of documents in the corpus where the word occurs. The Average Reduced Frequency (ARF) is a modified frequency whose calculation prevents the results from being influenced by a high concentration of a token in only one or more small parts (in my case, articles) of the corpus. If the token is evenly distributed in the corpus, ARF and absolute frequency have similar or identical values. The last column reports the Keyness scores for each word, calculated with the simple mathematical method described above (Equation 2). In the AER/Econometrica corpus, one of the first keywords is “intertemporal”. It has a Keyness score of 350.6, and it occurs in 28 different articles; we will analyze its collocations in the next section, in order to ascertain its centrality.

45In Table 1, I have selected, among the first 100 of the AER/Econometrica corpus, all keywords related to ‘technicalities’. By ‘technicalities’, I mean here, notably, words indicating mathematical properties of preferences or of the (representative) agents’ utility function (such as “quantile”, or “concave”), mathematical or statistical words (“corollary” and “covariance” for instance), or words related to individual choices (“multiperiod” and “maximisation”). In the Journal Economic Issues corpus, the keywords are not as technical as in the AER/Econometrica—the lexical choices are less specific. Among the first 100 keywords of the Journal of Economic Issues, there are actually none referred to ‘technicalities’, as defined above. This can be thought of as a negatively formulated finding.

46Another piece of evidence pointing at the higher technicalities of the articles published by the AER/Econometrica is the collocational comparison for the adjective predicates of “preference” (Figure 1).

48Figure 2 deals in the same way with the modifiers of “action”; and it shows that “action” has semantic domain of subjectivity for the Journal of Economic Issues and of objectivity for the AER. Indeed, in the former, there are psychological (“precipitous”), social (“collective”), and subjective connotations in general. In the latter, on the other hand, there are many objective connotations (“future”, “(in)feasible”, “set”, “optimal”). Some modifiers of “action” are shared by the two ‘subcorpora’; (“certain”, “policy”) and have similar relative frequencies.

49Other evidence of the technicality of neoclassical mold in the language used by the AER/Econometrica articles (with respect to those from Journal of Economic Issues), can be found in the relative frequencies of the lemma “proof”: 657.84 for the first, and 6.02 for the second. The relative densities are respectively 131.08% and 1,2%. A relative density above 100% means that is more frequent in the related ‘subcorpus’ than in the whole corpus. The same goes for the lemma “equation”.

3.2 Individual Needs or Social Process?

50The second result of our analysis is that the orthodox corpus mainly refers to the intertemporal aspect of preferences, to the absolute nature of consumption, to the representative agent in a maximization framework, and to the sphere of rational choice. But keywords and the semantic analysis of target lemmas reveals also the attention paid to the heterogeneity of preferences (although this concerns only one article) and to habitual behavior. The heterodox corpus discusses the relative (social) nature of consumption, the interdependence and interpersonal aspects of preferences, the concern for status, identity, and a variety of behaviors related to the social sphere.

51The empirical evidence sustaining this second result requires a deeper and longer analysis—therefore I will proceed in stages. First of all, I will present the evidence that shows that articles published in the AER/Econometrica place their emphasis on the autonomous aspect of consumption and preferences, while those from the Journal of Economic Issues place their emphasis on the social one. The identification of the semantic domain of the two corpora is performed through the analysis of keywords and multi-keywords. The latter are more likely to give a rather precise idea of the underlying theoretical differences.

Table 2. Keywords in the Semantic Sphere of Relative Consumption

		Frequency per million		Relative DOCF				ARF
Word		Focus	Reference	Focus		Reference		Focus	Reference	Score
1	duesenberry	405.23	< 0.01		21.31%	< 0.01%	29.57		59.13	404.9
2	conspicuous	1,113.38	2.78		62.30%	0.12%	160.53		68,504.88	294.9
3	institutionalist	170.52	0.04		36.07%	< 0.01%	29.47		1,003.17	164.2
4	positional	361.10	1.55		16.39%	0.05%	20.51		30,336.34	141.7
5	intersubjective	142.43	0.14		4.92%	< 0.01%	5.93		2,785.09	125.6
6	habitus	114.35	0.18		4.92%	< 0.01%	4.76		3,346.37	98.1
7	easterlin	94.29	0.03		9.84%	< 0.01%	8.85		450.71	92.8
8	overindebtedness	76.23	0.01		1.64%	< 0.01%	1.83		179.30	76.4

Keywords in the articles published by the Journal of Economic Issues. Keywords in this table are selected from the first 100 keywords.

52Table 2 reports the first 100 keywords of the heterodox corpus. All the words immediately refer to the social aspects of consumption (“positional”, “conspicuous”, “intersubjective”, “positional”) or to authors that, as we recalled supra (Section 1.1), provided cornerstone contributions to the relative hypothesis (“Easterlin”, “Duesenberry”). “Habitus” is the only word whose meaning may not be immediate to the (economic) reader: it refers to Bourdieu’s sociological concept. This word, compared to other keywords, is both an indicator of interdisciplinarity and of attention to the social aspects of consumption.

53Among the keywords of the Journal of Economic Issues corpus there is also the word “vicarious” (not reported in Table 2, since it is besides keyword number 100). This word indicates a certain degree of depth of some theories. Specifically, this word refers to the Veblenian concept of “vicarious consumption”, that is, consumption by proxy, through which an individual, maintained by or subordinated to/dependent from another one, shows the greatness of the latter (his/her master) through the conspicuous leisure or consumption. The ideas of Veblen usually entering the economic literature are the notions of conspicuous consumption, conspicuous leisure and status, but often referred to the single individual who signals his own status through these two forms of ostentation. By contrast, the concept of “vicarious consumption/leisure” refers even more the social character of consumption, in this case inextricably linked to a relationship between at least three individuals (for instance, a wife who does not work, but cares about showing to society the family’s wealth).

54A rather surprising result is that there is the relative absence of behavioral words in the publications about consumption in orthodox journals. At least from the AER, one would have expected more terms referring to the recent concepts of behavioral economics, such as, for instance, bounded rationality, status-driven behavior, agent myopia, equilibrium instability, out of equilibria dynamics, or errors/heuristics/biases in implementing consumption plans and others. Instead, just “inattentive consumer” is mentioned (in one article). The other only relatively behavioral element found in publications by orthodox journals is agents’ heterogeneity.

55Some relevant multi-keywords for identifying the semantic domain of the whole corpus are reported in the Online Appendix (Table A1). These multi-keywords tables provide further insights about the main semantic fields of both corpora, confirming the main finding outlined by Table A2 in the Online Appendix, related to the heterodox ‘subcorpus’.

56Three main features of the orthodox corpus are conveyed by Table 4 below: the temporal dimension of the mathematical problem, through lemmas such as “recursive”, “maximization”, “time-additive”, the intertemporal nature of agents’ choice (“intertemporal”, “one-period”, “multiperiod”, “life-cycle”) and the optimality/maximization as model’s trigger of the mathematical dynamic equations. In the AER/Econometrica corpus, we also find “martingale” and “recursive”, with a high Keyness score, revealing the influence of Hall (1978) on this corpus.

Table 3. Keywords in the Semantic Sphere of Intertemporality for AER/Econometrica

		Frequency per million		Document frequency			Relative DOCF			ARF
Word		Focus	Reference	Focus		Reference	Focus		Reference	Focus		Reference	Score
1	intertemporal	376.14	0.08	28	2,250		68.29%	< 0.01%		85.35	1,641.25		350.6
2	life-cycle	256.53	0.91	15	30,112		36.59%	0.04%		40.40	21,049.54		135.1
3	one-period	119.61	< 0.01	10	282		24.39%	< 0.01%		12.65	202.10		119.6
4	multiperiod	107.02	< 0.01	8	254		19.51%	< 0.01%		12.13	176.53		107.2
5	euler	193.58	0.84	14	14,764		34.15%	0.02%		33.59	11,413.46		105.5
6	optimal	1,262.18	13.58	36	442,057		87.80%	0.54%		301.50	308,977.75		86.6
7	recursive	180.99	1.34	16	33,257		39.02%	0.04%		31.73	25,074.49		77.7
8	maximization	86.56	0.44	19	14,799		46.34%	0.02%		25.37	10,320.80		60.7
9	time-additive	26.75	< 0.01	3	6		7.32%	< 0.01%		3.29	3.65		27.7

57The multi-keywords tables (see Online Appendix) sketch the broader semantic fields of both the corpora. The comment on these tables would be repetitive, with respect to the characteristics already highlighted so far, although they provide more precise and compelling evidence for the main findings of this article.

58Keywords collocations provide an important tool to deepen our understanding of the semantic usage of a word in a corpus. Figure 3 reports, for instance, the collocational behavior of the lemma “intertemporal” in the articles about consumption published by the AER/Econometrica.

59A collocational graph can be interpreted as follows. The higher the distance from the center, the lower the typicality score (i.e., the LogDice score, which represents the strength of a given link in a given corpus). The circle size indicates the absolute frequency of the collocation. Each color represents a different grammatical relation between the node and the collocate. The ‘slices’ sizes approximate the number of collocates in that specific grammatical relationship with the node.

60Figure 3 shows that “intertemporal” is often used as a modifier of nouns, and that it is semantically related to the agent preferences or utility function. In essence, Figure 3 shows that intertemporality is the main object of consumer choice. Comparatively, as we will see in Figure 5 (about the use of the word “preference”), the Journal of Economic Issues places social motivations at the center of consumer choice. The way of using the lemma “status” (that in the AER/Econometrica corpus occurs just once) is also indicative of the attitude of heterodox economists towards consumption theory, as reported by Figure 4.

61The variety of collocations is wide, and it relates back to TTR—the higher the TTR, the higher the variety and richness of language, and, thus, the higher the variety of collocations. The presence, among the modifiers listed in Figure 4, of the word “performative” suggests that, in the articles published by this journal, consumer theory is integrated with the theory of performativity (which links consumption and identity as production or performance of the self), at least partially (see e.g., Joseph, 1998). Furthermore, one can see that status is semantically collocated as something that agents aspire to and that conditions their choices.

62Figure 5 confirms that in the orthodox corpus, the temporal aspect of preferences and certain technical properties of preferences prevail (although one could notice that some attention is devoted to the heterogeneity of the preferences). Whereas, in the heterodox corpus, a view of preferences as socially determined as well as linked to a judgement of taste, prevails. The modifiers of “preference” are the purple slice of Figure 5, and they deserve a separate analysis in Figure 6.

63Among the shared collocations, “preference” and “consumer” are slightly stronger in the heterodox corpus, while “individual” and “time” are stronger in the orthodox one. The rest of the collocations are a closer focus on “preference” modifiers and confirm in detail the results in Figure 5.

64Comparing the way the corpora collocate the lemma “consumption” (Figure 7) is very insightful and strengthens the interpretation that articles in the AER/Econometrica are much more oriented towards the individual and intertemporal consumption, while those from the Journal of Economic Issues sees consumption as a social process. Moreover, the connotation of “consumption” collocations is much closer to the semantic sphere of objectivity in the orthodox corpus and of subjectivity in the other.

65The strongest collocations of “consumption” in the AER/Econometrica corpus are “good” and “smooth”. The latter appears both as a verb (in two grammatical relationships) and as a modifier. It still refers to a temporal dimension: the well-known result in economics of consumption smoothing, elaborated around the 1950s by Modigliani and Brumberg (1954; see Section 1). The articles published in the Journal of Economic Issues instead have “distribution” as most typical collocation and “conspicuous” as the more frequent one (and also a strong one, since it is very close to the center of the figure). The collocations of “consumption” with “relative” in the AER/Econometrica corpus is analyzed through the concordance: six occurrences of “consumption” with “relative” always refer to the future consumption compared to the past; and in two cases to the habit formation models. In a nutshell, they never refer to the social comparisons behind consumption, not even to the connotation of goods and services that make them act as social distinctive signs. In the Journal of Economic Issues corpus, the lemma “consumption” presents many collocations referring to the social influences on consumer choices. The strongest collocation in the “and”/“or” grammatical relation are “distribution” and “identity”.

66The comparison in the use of “behaviour” in Figure 8 is indicative of the greater determinism sought by mainstream theory than by heterodox theory.

67A higher determinism (meaning that the consumer’s behavioral rule is deterministic and hence the action is predictable) emerges from n-grams such as “rule of thumb” and “hand to mouth”, occurring in the AER/Econometrica corpus. Among the shared locations between the two corpora, we find “rational”. And yet, when checking the concordance of “rational” in the heterodox corpus, we find that it is used in sentences that critically discuss the rational agent framework—whereas in the orthodox corpus it is used simply as a qualifying adjective. For the lemma “consumer”, analyzed in Table 4 and 5, we observe something similar to Figure 7: the verbs using consumer as subject are “learn”, “perceive”, and “emulate” in the Journal of Economic Issues corpus, and “purchase”, “respond”, and “maximize” in the AER/Econometrica corpus.

Table 4. Word Sketch Differences for the Lemma “Consumer”: Verbs with Consumer as a Subject

	Relative Frequences (Journal of Economic Issues)	Relative Frequences (AER/Econometrica)	LogDice Score (Journal of Economic Issues)	LogDice Score (AER/Econometrica)
learn	7	0	9.5	-
want	7	0	9.5	-
perceive	5	0	9.1	-
pay	5	0	9.0	-
emulate	4	0	8.8	-
choose	6	10	9.0	10.0
face	4	8	8.5	9.8
purchase	3	14	8.3	10.8
die	0	4	-	9.1
react	0	5	-	9.5
visit	0	5	-	9.5
respond	0	6	-	9.6

Table 5. Word Sketch Differences for the Lemma “Consumer”: Modifiers of “Consumer”

	Relative Frequences (Journal of Economic Issues)	Relative Frequences (AER/Econometrica)	LogDice Score (Journal of Economic Issues)	LogDice Score (AER/Econometrica)
conspicuous	30	0	10.5	-
individual	13	0	9.6	-
household	6	0	7.8	-
many	3	4	7.6	8.1
more	0	5	-	7.9
few	0	3	-	8.5
marginal	0	7	-	8.9
wealthy	0	3	-	9.5
habit-forming	0	3	-	9.5
inattentive	0	3	-	9.5
heterogeneous	0	4	-	9.7
informed	0	4	-	9.9
hand-to-mouth	0	5	-	9.9
rule-of-thumb	0	8	-	10.8
representative	0	12	-	11.1

68The four columns of each table in Table 4 and 5 are, respectively: the occurrences per million of tokens (relative frequencies) in the heterodox corpus; the same measure in the orthodox one; the LogDice of the heterodox corpus; that one of the orthodox one. A blank space indicates there is no occurrence of that word in that corpus. The dichotomy between modifiers “conspicuous” and “representative” in the right table is emblematic of the differences between the two approaches.

69A surprising feature of these results is that the Journal of Economic Issues does not even mention the “representative consumer” to criticize this concept. Moreover, the higher determinism of the AER/Econometrica corpus is also found for “consumer”, as well as for “behavior”. The word “inattentive” also appears in Table 5, indicating that even orthodox theory has moved closer to the concept of bounded rationality, although this occurrence comes from only one article.

Table 6. Word Sketch Differences for the Lemma “Consumer”: Nouns Modified by “Consumer”

	Relative Frequences (Journal of Economic Issues)	Relative Frequences (AER/Econometrica)	LogDice Score (Journal of Economic Issues)	LogDice Score (AER/Econometrica)
boycott	49	0	10.7	-
theory	38	0	9.4	-
demand	65	9	10.8	8.9
debt	63	12	10.5	9.0
spending	34	6	10.0	8.8
behavior	108	40	10.9	10.1
choice	34	11	9.9	9.4
credit	78	31	11.1	10.9
finance	33	11	10.1	10.0
durable	0	4	-	8.8
surplus	0	17	-	10.6
search	0	21	-	11.1

Table 7. Word Sketch Differences for the Lemma “Consumer”: Nouns Modified by “Consumer’s”

	Relative Frequences (Journal of Economic Issues)	Relative Frequences (AER/Econometrica)	LogDice Score (Journal of Economic Issues)	LogDice Score (AER/Econometrica)
rate	6	0	10.9	-
taste	4	0	10.4	-
repair	3	0	10.1	-
decision-making	3	0	10.1	-
want	3	0	10.0	-
preference	7	5	10.8	10.2
behavior	4	8	10.0	10.9
decision	3	8	9.7	11.0
wealth	0	5	-	10.4
value	0	5	-	10.5
utility	0	10	-	11.4
problem	0	13	-	11.5

70Tables 6 and 7 complements Table 4 and 5 and show that the shared collocations are few and generic, two of which relate to the financial sphere. Also “expectation” (Figure 9) is used in a way that confirms the formalism of the articles published by the AER/Econometrica in contrast with those published in the Journal of Economic Issues.

71“Reciprocal”, in Figure 9, refers to Habermas’ concept of reciprocal expectation, that is, the idea that social knowledge is governed by binding consensual norms.

72This section showed that orthodox consumer theory is much more focused on the individual and his choices based mainly on inter-temporal factors (i.e. dependent on himself and objective market conditions), while heterodox theory pays attention to the social determinations of consumption choices. Aspects related to ‘technicalities’ (see Section 3.1) can also be found again here.

3.3 Theoretical Debate and Interdisciplinary Approach

73The third result of our analysis, that is, the more substantial engagement of heterodox articles with theoretical debates, comes from observing the usage of some selected lemmas (such as “theory”, “approach”, and “hypothesis”) as well as the usage of authors’ personal names.

74The first striking difference between our two corpora is that the word “theory” appears much more in the heterodox corpus than in the orthodox one: relative frequencies for this word are, respectively, 1,947.07 and 385.08. This is also the case for the word “approach” (relative frequency 481.46 in the heterodox corpus and 262.82 in the orthodox one). The opposite finding holds for “hypothesis” (relative frequence 355.68 in the in the orthodox corpus and 262.80 in the heterodox one).

75Verbs with “theory” as an object or subject semantically refer to a discussion or hypothesis in the Journal of Economic Issues corpus (“reconsider”, “reject”, “propose” and “suggest”), while they are related to more decisive statements in the AER/Econometrica corpus (“expect”, “imply”, “predict”, and “enable”). The same is true for “hypothesis” (see Figure 10), whose collocation patterns are strongly related to verbs as “reject”, “satisfy”, “contradict”, and “fail” in the orthodox corpus, and to “reposition”, “defend”, “found”, “ground”, “explore”, and “examine” in the heterodox one. From this, we see the tendency to use verbs semantically linked to a form of determinism.

76From the modifiers and the possessors of these lemmas we understand how the two corpora differ among them within their way of discussing economic theories.

77Within economic theory, the analysis of collocations of “hypothesis” synthesizes the theories and hypotheses of the two corpora: Friedman’s Permanent income, Modigliani’s life cycle, and Hall’s rational expectation are the most represented hypothesis in the articles published by the AER/Econometrica; articles in the Journal of Economic Issues instead refer mostly to Minsky’s financial instability hypothesis, to Duesenberry’s relative income hypothesis, and to the Veblenian concept of conspicuous consumption.

78The word “theory” in the heterodox corpus is more often associated (in conjunction or with its possessors) with authors’ personal names (“Veblen”, “Bourdieu”, “Galbraith”, “Foley”, and “Keynes”).

79Each specification of the word “theory” refers to a precise theory or author; the fact reflected in the data is that the Journal of Economic Issues has a broader theoretical spectrum than the AER and Econometrica in terms of frequencies of author personal names, as it was suggested by the keywords extraction. This leads us to the fourth result of our analysis, that is, that the heterodox corpus engages more substantially with non-economic literature. For instance, theories mentioned in this corpus come from a greater variety of disciplinary fields. Looking at “theory of” specification in the orthodox corpus, the main disciplinary fields involved are economics and statistics; while, for the heterodox corpus, the theories come from various non-economic fields, including sociology (“theory of action” by Talcott Parsons), philosophy (“theory of mind” by John Locke), psychology (“theory of cognitive dissonance” by Leon Festinger), anthropology (“theory of cultural evolution”) and marketing (“theory of consumer behavior”).

80Figure 11 illustrates that the main semantic clusters suggested by the modifiers consist either, on the one hand, in attributes that presuppose their opposite (or, at least, the choice between two or more theories), or, on the other hand, in attributes that do not contemplate such possibility. Among the words in green, the attributes of “theory” (“conventional”, “mainstream”, “classic”, “traditional”, “alternative”, “orthodox”, “different” and “comparison”) embody, within their semantics, a comparative stance towards alternative theoretical visions of consumption. Among the modifiers of “theory” in the orthodox corpus, there are no attributes involving a comparison among different theories. The figure of the modifiers of “approach” (Figure 12) is also interesting in order to understand the patterns of the lemma by the two corpora.

81“Computational”, “non parametric”, “structural”, and “control function” show that articles published by the AER/Econometrica use the lemma “approach” for discerning between different (mathematical) methods for solving a problem, while publications in the Journal of Economic Issues use the word “approach” for distinguishing points of view.

82The interdisciplinarity of the articles about consumption published by the Journal of Economic Issues (with respect to those of the AER/Econometrica corpus) is revealed above all by the keywords. The Journal of Economic Issues keywords’ table (see Table 8 below) presents a lot of authors personal names. These authors belong to different disciplines (Bordieu to sociology, Polany to anthropology, Vas Verplanken to social psychology), research areas (Minsky to financial and monetary economics, Hirschman to development economics), and to different approaches (Kenneth Boulding to the evolutionary economics). Other authors entering the heterodox keywords are Easterlin (cf. supra, 3.1) and Albert Bandura (a psychologist).

83The first ten authors personal names in both the corpora are reported below: for the Journal of Economic Issues, they are among the first 15 keywords; for the AER/Econometrica corpus, they are among the first 198; and this is already a result in itself: heterodox corpus pays much more attention to the milestones. One could venture an interpretation: heterodox theory always starts from cornerstone authors (from disparate fields), while the mainstream corpus builds up theory more spirally, adding piece by piece and losing sight of the theoretical underpinning. But this is perhaps a far-fetched interpretation.

Table 8. First Ten Authors among the Journal of Economic Issues Keywords

		Frequency per million		Relative DOCF		ARF
Word		Focus	Reference	Focus	Reference	Focus	Reference	Score
1	Veblen	2,591.87	0.17	75.41%	< 0.01%	432.87	2,525.91	2,219.9
2	Bourdieu	601.83	0.29	16.39%	< 0.01%	31.62	4,405.71	467.3
3	Duesenberry	405.23	<0.01	21.31%	< 0.01%	29.57	59.13	404.9
4	Minsky	198.60	0.25	14.75%	< 0.01%	13.40	4,228.49	159.6
5	Bentham	196.60	0.54	16.39%	0.01%	15.88	8,796.93	128.2
6	Polanyi	134.41	0.18	8.20%	< 0.01%	8.08	2,588.47	115.0
7	Boulding	122.37	0.07	3.28%	< 0.01%	4.52	1,319.23	114.8
8	Easterlin	94.29	0.03	9.84%	< 0.01%	8.85	450.71	92.8
9	Bandura	94.29	0.15	1.64%	< 0.01%	1.75	2,336.59	83.1
10	Scitovsky	80.24	<0.01	13.11%	< 0.01%	11.37	103.18	80.7

84Among the first ten authors’ personal names in the Journal of Economic Issues corpus there is a sociologist (Bourdieu), a philosopher (Bentham), an anthropologist (Polanyi) and a psychologist (Bandura).

Table 9. First Ten Authors among the AER/Econometrica Keywords

		Frequency per million			Document frequency				Relative DOCF			ARF
Word		Focus		Reference	Focus		Reference		Focus		Reference	Focus		Reference	Score
1	Euler	193.58	0.84		14	14,764		34.15%		0.02%		33.59	11,413.46		105.5
2	Pesendorfer	59.80	< 0.01		2	115		4.88%		< 0.01%		3.78	79.48		60.6
3	Attanasio	61.38	0.05		7	1,375		17.07%		< 0.01%		13.11	981.79		59.5
4	Galvao	56.66	0.03		1	577		2.44%		< 0.01%		2.82	457.44		56.0
5	Zeldes	55.08	0.01		8	298		19.51%		< 0.01%		12.17	223.79		55.5
6	Markov	97.57	0.91		10	21,762		24.39%		0.03%		16.69	16,571.61		51.5
7	Nakajima	64.53	0.30		1	6,712		2.44%		< 0.01%		3.30	4,529.48		50.4
8	Eichenbaum	50.36	0.02		6	2,179		7.32%		< 0.01%		9.64	416.54		50.4
9	Newey	53.51	0.09		3	13,885		2.44%		< 0.01%		5.46	1,626.00		49.9
10	Chatterjee	72.39	0.51		1	576		14.63%		0.02%		3.58	9,424.33		48.6

85In the orthodox corpus all the cited authors are economists, excluding two cases—two mathematicians (Euler and Markov) whose name is rather mentioned because it is attached to a specific technique or concept (Euler equations and Markov chains).

86These final set of results show that articles about consumption published in the Journal of Economic Issues discuss the neoclassical approach, rationality, and other concepts. This corpus employs a language conveying a general attitude of openness to discussion with the other stream of literature on consumption. But there seems to be no one to debate with, since, on the other hand, the AER/Econometrica corpus on consumption analyzed here do not bother to mention neither the contributors to their own approach (neoclassical), nor others (institutionalist, evolutionist or behaviorist for instance). The AER/Econometrica corpus does not talk about “approaches”, it does refer to authors active strictly within the field of economics—in short, it is more self-referential. The heterodox corpus talks about approaches continuously. It’s indicative in this regard, the fact that in the heterodox corpus the word “neoclassical” is the tenth keyword, with 154.8 score, while in the orthodox corpus is not even among the first 500 keywords even if it is its own theoretical underpinning. The same remark applies to the word “rationality”, much more discussed by articles in the Journal of Economic Issues than by those published in the other two journals (4 occurrences in the AER/Econometrica, and 74 in the Journal of Economic Issues), even if they use it as hypothesis. The opposite is true for “rational”: 244 (267.79) occurrences in the AER/Econometrica and 76 (173.28) in the Journal of Economic Issues (relative frequencies per million of tokens are in parenthesis).

87Other evidence of absence of debate in the AER/Econometrica are the words “heterodox” and “orthodox”, which in the Journal of Economic Issues are contemplated while in the other two journals they never appear. Given these results, the debate between the heterodox and the orthodox economics is likely to be sterile, because in the end, despite all the efforts of a faction, the intention to debate remains one-sided.

Conclusion

88This article shows that computational linguistics can help us to identify theoretical standpoints and theoretical underpinning of a corpus of academic articles. In our case study about consumption theories, we have been able to analyze two corpora, to show that they linguistically differ, and that this difference can be attributed to different theoretical views.

89The Journal of Economic Issues publishes articles on consumption with a higher number of references to authors from the most disparate research fields. These articles also exhibit a greater lexical variety and they are less technical than those published by the AER/Econometrica; Journal of Economic Issues articles use less mathematical symbols (if we take tokens with respect to words ratio to be a good proxy), and they devote greater attention to the interdependent, sociological, and psychological aspects of preferences. In the AER/Econometrica corpus, attention is instead focused on the intertemporal aspects of consumer choices in a maximizing framework.

90Text mining can come to the aid of the history of economic thought especially for large datasets. The dataset used here is not large, constrained by the computational possibilities available when the research was carried out. Yet, 112 articles is still a number that justifies the use of computational software. The big difference between the computational linguistics techniques used in this work and other techniques (see for instance Oliveira and Dávila-Fernández, 2020; Tusset, 2021; Kufenko and Geiger, 2016), which make use of the ‘bags of word’ approach) is that with computational linguistics you keep a control over the meaning of the original text; we may say that the techniques exploited here are a bridge between pure quantitative analysis and the traditional exegesis of the text. The metrics used in this article are only two scoring functions (the Keyness score and the LogDice score), but they can be employed for a variety of purposes. Here, too, words are transformed into numbers, but you can always verify how the specific word is used in the specific sentence.

91This linguistic technique proved to be useful to extract patterns of attitude with respect to a particular topic of different groups of authors; in Section 3.3, it is also used to test group’s attitude towards pure theoretical debate, and to measure their level of interdisciplinarity. As for this last point, simply mapping the relative frequencies of the cited authors, we document which names are most relevant (if frequency is a good proxy of relevance) in a given journal on the topic of consumption. But the interdisciplinarity may be measured also from the lexical choices: “habitus” is an example of word drawn explicitly from the sociological field.

92Let us now go back to the theoretical discussion (related to Section 1) on consumption that follows the observed empirical findings. It is difficult to enclose preference interdependence in the analysis of demand formation; for two reasons, a technical one and an ideological one. The technical one concerns the difficulty raised by the preference interdependence of obtaining aggregate demand as the sum of individual ones (Drakopoulos, 2012). The ideological reason is that admitting the social nature of consumer choices challenges the idea of consumer’s sovereignty, a cornerstone of liberal ideals.

93The differences in theorizing consumption are not politically neutral, they have different consequences both in fiscal (Drakopoulos (2021a), Samuelson (2004), Aronsson and Johansson-Stenman (2008)) and environmental terms (Howarth (2006), Shove and Warde (2002)). About the fiscal consequences there are several implications, but the basic insight is that under preference interdependence a redistributive fiscal policy is Pareto efficient, as mentioned above. About environmental consequences instead: “Accounting for relative consumption effects reduces the perceived social benefits of consumption, thus increasing social willingness to pay for environmental quality while reducing both the costs of tax interaction effects and the benefits of efficient revenue recycling. Taken together, these factors support the implementation of comparatively high environmental taxes in the face of relative consumption effects.” (Howarth, 2006, 1)

94This article proposes an exploration of recent history of economic thought through computational linguistic methods. It has limited explanatory power about the approach to consumption of economics as a whole, since results are limited to the journals considered. Text mining should be used with caution and certainly cannot replace qualitative analysis, but, at most, they complement it.

I would like to express my sincere gratitude to the two referees for their valuable comments and suggestions, which greatly improved this work, and to Francesco Sergi for his support and insightful discussions.