Historiska persistensstudier

Inom ekonomisk historia, särskilt så som den praktiseras inom nationalekonomin, har studier av historisk persistens slagit igenom stort de senaste tjugo åren: medeltida pogromer bestämmer naziströstning på 1930-talet, plogens genomslag eller inte på stenåldern bestämmer variationer i jämställdhet mellan könen idag, digerdöden bestämmer det mesta, och så vidare. I senaste utgåvan av Annual Review of Political Science konstaterar statsvetarna Alexandra Cirone och Thomas B. Pepinsky att även i deras disciplin har persistensstudierna slagit igenom*, inspirerat av "the rise of quantitative approaches in economic history". De definierar persistens så här:

"we define historical persistence as causal effects that (a) operate over time scales of a decade or more and (b) explain spatial variation in political, economic, or social outcomes. The decision to set a decade-long time scale for long-term persistence is arbitrary, but it distinguishes historical persistence from analyses that explore short-run political dynamics or structural causal relationships without a historical component. The emphasis on spatial variation in outcomes, in turn, distinguishes historical persistence from general claims about the causal weight of history, such as the observation that Portuguese colonialism caused contemporary Brazil to be a Lusophone country. Our analysis of historical persistence sees the central analytical task as estimating the effect of spatial variables that operate over a substantial temporal distance." (s. 241)

Persistens-studierna använder kvantitativa metoder, och vill urskilja kausala effekter.

Persistens-vågen började med två "landmark essays" av Daron Acemoglu, Simon Johnson och James Robinson: AJR 2001 och 2002. De publicerades inte i statsvetenskapliga tidskrifter, men "these two essays quickly became standard reading among political economists, owing to their powerful demonstration of how historical data could be assembled to produce causal explanations for the contemporary distribution of political institutions across the world" (s. 242) Sokoloff och Engerman (2000) och andra hade gjort kvantitativa studier av kolonialismens långsiktiga politiska effekter, men vad som var unikt med AJR var att de kombinerade historiska data med " a framework for causal identification to estimate effects that operate at century-long time scales." Detta tillvägagångssätt blev centralt för -- definierar -- persistens-studierna. 

Cirone och Pepinsky har gått igenom fem nationalekonomiska och nio statsvetenskapliga tidskrifter och kodat artiklarna utifrån abstract till om det handlar om persistens eller ej. Inom nationalekonomi gick andelen persistens-studier från obefintlig till 5-10 procent vid 2010-talets början; inom statsvetenskap gick andelen över fem procent först 2016-17 och har sedan dess pendlat mellan 5 och 20 procent varje år. Studierna handlar om en mängd olika saker, men Cirone och Pepinsky identifierar ett par kluster.

Ett kluster följer direkt på AJR (2001, 2002) och handlar om kolonialismens långsiktiga effekter. Ett relaturat kluster handlar om effekterna av prekoloniala statsformer. Dessa sudier handlar oftast om globala Syd; ett undantag är Voglers (2019) studie av Polen. Ett tredje kluster handlar mer om globala Nord och handlar om ursprunget till och effekter av Förintelsen i Europa, och andra former av våld och masspolitik. Ett fjärde kluster handlar om "agglomeration and the infrastructural foundations of contemporary economic development." Ett femte kluster handlar om slaveriets långsiktiga effekter, i USA och Nord såväl som i Afrika söder om Sahara.

Vad gäller tidsskala är de mest långsiktiga persistensstudier de som handlar om ett par årtusenden, som Alesina et al (2011) om plogen under stenåldern. Nationalekonomerna skriver också i genomsnitt persistensstudier med längre tidsperspektiv än vad statsvetarna gör. Geografiskt sett så är spridningen ganska jämn: det finns många studier om Europa och Nordamerika, men också ganska många om Afrika söder om Sahara och Asien. Däremot handlar ganska få studier om Sydamerika eller Mellanöstern.

Efter beskrivningarna av persistensstudierna, så övergår Cirone och Pepinsky till en fråga som jag tycker är hyperintressant: hur och varför kan historiska orsaker ha bestående effekter? De typologiserar dessa förklaringar i två typer: equilibrium dependence och slow change. Så här förklarar de den första typen:

"Equilibrium dependence encompasses a set of causal processes in which some prior characteristic of a system affects its long-term equilibrium state. Such processes are commonly (if erroneously) lumped together under the terms path dependence or increasing returns, but Page (2006) distinguishes formally among concepts such as increasing returns, self-reinforcement, positive feedbacks, and lock-in. The distinguishing feature of equilibrium dependence is the existence of more than one equilibrium distribution of outcomes, with the equilibrium depending on past realized outcomes.
Importantly, however, it could be that in the long run there exists only one single distribution of outcomes, but that the outcomes at any particular time t nevertheless depend on previous outcomes. Page (2006) terms this outcome dependence. Equilibrium dependence implies outcome dependence, but the reverse is not true. Outcome dependence produces long-term persistence of historical causes on subsequent outcomes due to the slow rate of change of initial conditions over time." (s. 245)

De förtydligar med hjälp av ett exempel där en jordbävning förstör en stad där husen oftast var vita. Åren efter jordbävningen är den billigaste husfärgen röd. Hur kommer korrelationen mellan husfärgerna året före jordbävningen, och husfärgerna hundra år senare, se ut? Ett annat exempel är en kolonial policy som gör att det finns missionärer i en del koloniserade distrikt men inte andra. Hur kommer detta påverka läskunnigheten hundra år senare, givet olika antaganden om skolbyggande etc? Dessa slutsatser drar Cirone och Pepinsky av den illustrativa diskussionen:

"The upshots of this discussion are twofold. First, many conceptual tools can explain theoretically why initial conditions may shape subsequent outcomes. A theoretically satisfying account of why different initial causes produce long-term variation in outcomes requires that the mechanisms of persistence be precisely characterized to understand why causal effects exist. Second, equilibrium dependence is not necessary for historical persistence. For many political science questions that examine persistence over decade-long time scales—relatively common among political science articles—outcome dependence may suffice to explain why spatial variation in outcomes persists. Political scientists need not cast their findings in terms of path dependence or something like it to identify causal effects over time. Equilibrium dependence is likely more important for explaining very long-term persistence, over century-long time scales." (s. 247)

Från detta övergår de till en metoddiskussion om de kvantitativa metoder som forskarna i persistensstudier använt.

"Abramson & Boix (2019) use data on geographic and climatic features, urban population, industrialization, income, and early representative institutions, for both 225 × 225-km grid-square units and autonomous political units across Europe, from 1200 to 1900, to test competing theories of the origins of political–economic divergence. The authors utilize such an expansive time series and cross-sectional data because they argue that prior studies’ data probably begin after divergence took place. Similar data collection efforts are used to study the development of the modern state, from military capabilities to democratic institutions (e.g., Blaydes & Paik 2016, Cox & Dincecco 2021, Stasavage 2020)." (s. 248)

Cross-section-time-series-data har också använts t ex av Suryanarayan (2018) som använder mikrodata från Indien för att utforska under vilka förhållanden fattiga blir mer benägna att rösta på högerpartier. Hon använder folkräkningsdata från 1931, som kodade kasttillhöirghet, för att karaktärisera dagens politiska distrikt och regioner. I USA använder Bazzi et al (2020) 150 år av folkräkningsdata för att geokoda hur olika delar av USA fick olika politiska kulturer beroende på om de var på "the frontier" eller inte. Så här diskuterar Cirone och Pepinsky nackdelarna med en CSTS-approach:

"The challenge of studying historical persistence using standard regression analyses comes
from omitted variable bias or endogeneity, even when one controls for unit or time fixed effects (Blackwell & Glynn 2018). Since this approach is observational, and depends on ex post statistical adjustment, it requires a strong research design to defend it. As a result, the best of such studies demonstrate clear, theoretically grounded hypotheses, comprehensive data, extensive discussion and testing of alternative hypotheses, and results that are robust to alternative model specifications." (s. 248)

Nästa approach som de diskuterar är naturliga experiment. "Natural experiments are observational studies where the causal variable varies as a result of natural, social, or political forces, and so the researcher can make a credible claim that treatment assignment is as-if random. Sometimes this involves exploiting historical lotteries that used an actual randomization device, such as in the case of committees during the French Third Republic (Cirone & Van Coppenolle 2019) or guild selection in medieval Florence (Abramson 2021)." (s. 248) Oftare handlar det inte om faktiska lotterier utan att forskaren får göra ett detaljerat argument för att folk och platser inte själva kunde välja om de skulle hamna i treatment-gruppen eller kontrollgruppen. Som exempel tar de Charnysh _(2019) som undersöker hur etnisk diveristet påverkat ekonomin i Polen efter andra världskriget. När gränserna drogs om mellan Tyskland och Polen var det med ett inslag av godtydlighet och slump; folk som berördes kunde inte själva välja var de skulle hamna. Så här avrundar C och P diskussionen om approachensför- och nackdelar:

"Claims to plausibly random treatment assignment can be defended using statistical tests that show balance between treatment and control groups, placebo analyses, or other sensitivity tests (Angrist & Pischke 2010, Cinelli & Hazlett 2020). But they also require qualitative, case-based arguments explaining why the treatment is probabilistic and unconfounded. This requires detailed knowledge of the historical case and efforts to avoid reading history backward—in other words, incorrectly imputing the causes of certain developments from their correlates or consequences (Capoccia & Ziblatt 2010, Møller 2021). As Kocher & Monteiro (2016, p. 952) write, “Qualitative historical knowledge is essential for validating natural experiments.”" (s. 249)

Den tredje approachen är vad de kallar "design-based inference". "Design-based inference exploits features of the research design—rather than model-based statistical adjustment—to move from correlation to causation (Sekhon 2009, Dunning 2012, Cantoni & Yuchtman 2020). Historical persistence studies commonly rely on three statistical techniques to study the past: RD designs, DiD designs, and IV analyses." Regression discontinuity har använts i en mängd kontexter:

"to look at the long-run effects of medieval or colonial institutions on state development in settings such as China (Mattingly 2015), Namibia (Lechler & McNamee 2018), Peru (Dell 2010), Spain (Oto-Peralías & Romero-Ávila 2017), Vietnam (Dell et al. 2018), and Kenya ( Jedwab et al. 2017). This method has also recently been used to examine how borders affect modern political outcomes, such as voting, extremism, or representation (Fontana et al. 2017, Fowler & Hall 2017, Tur-Prats & Valencia Caicedo 2020). Importantly, any RD design—historical or not—has high internal validity, but may have low external validity, for the sample around the threshold for which a causal effect may be estimated may be quite small or nonrepresentative. Further, this research design requires a significant number of observations around this threshold to estimate an effect, which might be challenging with historical data." (s. 249)

Difference-in-differences utgår från en händelse eller chock som ledde till senare olikheter mellan enheter som före chocken hade liknande trender. Cantoni et al (2018) utforskar hur katolska och protestantiska regioner utvecklades efter reformationen; Fouka (2020) studerar tvångsmässig assmileringspolitik med utgångspunkt i USA:s förbud mot tyskspråkiga skolor år 1919.

Instrumentvariabeldesignen "rely on a historical or geographic variable that serves as an instrument for
an endogenous regressor. In other words, this instrument is correlated with the treatment and
uncorrelated with the error term in the outcome model. This assumption is termed the exclusion
restriction. If it (among other conditions) holds, instrumental variables can allow scholars to isolate
the effects of historical variables on subsequent outcomes." (s. 250) Nunn (2008) och Nunn och Wantchekon (2011) analyserade effekterna av slavhandeln på Afrikas ekonomiska utveckling med avståndet till stora slavhandelshamnar som instrument för slavhandelns intensitet. Acharya et al (2018) visar att slaveriets intensitet i distrikt i USA år 1860 korrelerar med politiska attityder idag, och använder jordens lämplighet för att odla bomull som instrument för slavhandelns intensitet. Det finns också än mer exotiska användningar:

"Natural phenomena may also serve as instruments. Wang (2021) studies the effects of state-sponsored violence in 1960s China on modern political attitudes, using exposure to sulfur as an instrument: Proximity to sulfur mines resulted in arms manufacturing plants, which were well guarded by the state and suffered less violence than other regions. Sellars & Alix-Garcia (2018) explore how a demographic collapse in the seventeenth century concentrated land ownership in Mexico, which helps explain nineteenth- and twentieth-century landholding inequality and land reform. They use precipitation data (extracted from tree-ring chronologies) to instrument for disease prevalence and therefore variation in population levels." (s. 250)

Bland problemen för IV-approchen finns hur man ska övertyga om att the exclusion criterion verkligen funkar.

Efter denna metod-diskussion övergår C och P till de analytiska utmaningarna i persistens-litteraturen. Det handlar om fyra utmaningar: mått, spatial heterogenitet, posttreatment and collider bias, och kausala mekanismer. Vad gäller måtten är ett väldigt intressant problem detta: för att kunna studera persistens måste man ju ha ett historiskt mått på den oberoende variabeln: men vilka historiska data har egentligen överlevt in i vår egen tid? Där finns antagligen en tendens, t ex att undertryckta gruppers information i lägre grad bevarats. Frågan om var måtten kommer ifrån skulle jag t ex mena kan ställa till problem för Giuliano och Nunns mått på historisk bydemokrati, som ju uttryckligen ibland kommer från 1800-talet, ibland från 1900-talet, osv.

Frågan om spatial variation: persistens-studierna jämför distrikt, regioner och stater.

"But the spatial nature of the data means that adjacent observations may be dependent on each other (Cook et al. 2020). Correlations could potentially be driven by extreme levels of spatial autocorrelation of residuals or strong spatial trends, leading to biased coefficients (Casey & Klemp 2021) or artificially low standard errors (Kelly 2019). Kelly (2019) demonstrates this problem by assessing the extent of serial correlation in 27 persistence studies in top economics and political economy journals, and showing that if spatial noise is taken into account, many studies’ findings are no longer statistically significant. Yet as Voth (2021) highlights, many of these studies employ multiple specifications, and the results from the more sophisticated models that account for spatial dependence remain robust." (s. 251-2)

Geografiska fixed effects är en standard-metod som används; klustrade standardfel m m används också.

Den tredje utmaningen är posttreatment bias: om decennier eller århundraden skiljer oberoende och beroende variabel åt, vad kan inte ha hänt däremellan?

"... this means studies often include posttreatment variables: variables that occur after the proposed treatment has happened and lie on the causal path from the treatment of interest to the outcome. Conditioning on such variables can introduce posttreatment bias. Posttreatment bias can be introduced both by controlling for posttreatment variables in statistical models and by selecting the sample or subsetting the data on posttreatment criteria (Montgomery et al. 2018). Posttreatment data are also increasingly and incorrectly used to justify the exclusion restriction in IV analysis, which can actually undo all the benefits of a natural experiment. More generally, when working with missing, noisy, or aggregated historical data, it may be more challenging to accurately determine the exact timing of covariates, and to what extent one was measured before the other: Two variables may be recorded as coming from the seventeenth century, but if temporal order matters, such a coarse measurement may be insufficient.
If scholars do not adequately control for confounders, their analysis suffers from omitted variable bias, but what appear to be confounders might actually be causal consequences of the treatment variable. What are the solutions? First, these trade-offs can be modeled by using DAGs (directed acyclic graphs), a nonparametric approach to test whether causal effects are identifiable (Pearl 2009). Causal mediation analysis can also be used to examine intermediate variables that exist on the causal path between treatment and outcome (Imai et al. 2011), although the identification assumptions required for such analyses are strong. Acharya et al. (2016) propose the use of a two-stage regression estimator, the sequential g-estimator, to estimate the controlled direct effect, i.e., the treatment effect holding fixed values of a potential mediator. This has been employed in research on historical persistence (Charnysh 2019, Homola et al. 2020, Ito 2021), but its use still requires a properly specified causal model." (s. 252)

Vad gäller kausala mekanismer blir jag lite förvånad över hur avslappnade Cirone och Pepinsky är inför kausala förklaringar som inte kan ange mekanismerna genom vilka X påverkar Y: jag kan inte hålla med dem när de skriver att

"Under the standard approaches to causality in the quantitative social sciences (Rubin 2005, Pearl 2009), causal effects can be identified and estimated without reference to mechanisms. Under the potential outcomes framework for causality and many counterfactual-type definitions related to it, an explication of causal mechanisms is not necessary to estimate causal effects. The failure to supply evidence of causal mechanisms that link, for example, medieval anti-Semitism to Nazi-era violence against Jews (Voigtländer & Voth 2012) does not comprise evidence that this link is not causal." (s. 253)

Det låter obegripligt för mig. Efter ett sofistikerat resonemang om de kausala mekanismernas roll landar de i vilket fall i denna rekommendation för fortsatta persistensstudier:

"Although we welcome the discussion of causal mechanisms in historical persistence research, calls for mechanisms face the same challenges in the study of long-run persistence as they do in the rest of the social sciences. Even though historical persistence studies seem particularly in need of mechanismic evidence given the long time scales of the causal relations that they seek to identify, they are unlikely to prove dispositive for establishing that long-term correlations are indeed causal without an abundance of intermediate outcome data and multiple qualitative studies that meet stringent case selection criteria. Taking the causal ambitions of the historical persistence literature seriously means that the standards for mechanismic evidence to complement correlational
evidence are beyond the reach of most feasible studies.
We emphasize, however, that this is not an argument against mechanismic evidence. Like others, we expect that a middle ground is often possible, with historical and process-based evidence complementing the standard template, much as Simpser et al. (2018) have proposed for reconciling political economy approaches with comparative historical approaches to historical legacies.
Instead, we conclude that unconditional demands for mechanismic evidence are inappropriate for historical persistence studies and that the requirement that authors identify a single mechanism to the exclusion of all others is almost always unjustified. We recommend that quantitative scholars adopt a view of historical evidence that recognizes the limitations of mechanismic evidence in the context of their explanatory objectives." (s. 254)

Slutsatserna känns över huvud taget ganska preliminära: det är en litteratur som utvecklas snabbt och det blir ju också då svårare att sammanfatta vad den egentligen kommit fram till. I slutsatssektionen säger Cirone och Pepinsky att:

"As Nunn (2020) notes, for example, replicability is an issue with historical work. Data constraints in historical research make independent replicability across different empirical contexts difficult. But to the extent that replication is possible, it will strengthen the appeal of otherwise unique historical findings, moving beyond causal estimates to generalizability and event prediction.

Research on historical persistence can continue to theorize and model the passage of time (Grzymala-Busse 2010). Key questions include how to subdivide stretches of time when studying endogenous processes, what the mechanisms of persistence are, and what sorts of causal claims require affirmative mechanismic evidence. That physical infrastructure like roads and dams persists ever time is relatively uncontroversial, but just how social formations, norms, and ideas can persist over decades and centuries should remain an active area of research." (s. 254-255)

fotnoter

* det är inte heller den första översikten över persistensstudierna. Hans-Joachim Voth skrev 2020 "Persistence -- myth and mystery" (pdf) och Leticia Arroyo Abad och Noel Maurer publicerade 2021 "History Never Really Says Goodbye: A Critical Review of the Persistence Literature" i Journal of Historical Political Economy.

referens

Alexandra Cirone och Thomas B. Pepinsky, "Historical persistence", Annual Review of Political Science 25: 241-259, 2022.