This paper is my final project in my introduction stats class

Abstract

This study focuses on the relationship between descriptive social norms and pro-environmental behavior. I evaluate visual cues in the form of the fullness of a compost bin as evidence of others recycling behavior. I measured pro-environmental behavior in the form of an individual’s decision to sort their trash. Results for OLS regression analysis indicate that on average *ceteris paribus *every 10% increase in fullness of the bin leads to a 0.4% percentage point increase in likelihood to sort trash, these results are significant at the 5% level. Even a 1% change could diverts pounds from going to landfill daily. The main findings of this study imply that social norms can motivate people’s behavior without intrinsic motivation.

J**EL Classification**: D81, Q51

- Introduction

Recycling and composting are important for reducing landfill waste and creating a waste system. Motivating people to sort their recycling is difficult because an individual rarely benefits directly from sorting their own waste. Of course there are environmental benefits to recycling and composting properly, but those benefits are long-term and intangible. For these reasons, people will opt to throw all their waste in the landfill rather than sort.

Despite the availability of recycling and compost, sorting behavior rates are not as high a they can be. Sorting takes effort and time to do properly, and for many people the moral obligation or personal interest in recycling or compost does not overcome these felt costs of time and energy. Therefore, intervention is required to raise these rates. This paper focuses on cost-effective ways to increase sorting and sustainable disposal rates. More specifically, this paper tests the effect of descriptive norms. Descriptive norms are one’s perception of what others are doing and what is socially accepted as the status quo.

The question I will answer in this research essay is what effect do descriptive norms have on someone’s likelihood to sort their trash. First, in section II, I will contextualize my topic within the context of environmental social studies and behavioral economics. In section III, I will describe my multiple regression model. Then, in section IV, I will present my methods for data collection. Finally, in Section V I offer results from the study and discussion of my findings. Lastly, Section V provides concluding remarks on the significance of my data.

- Lit Review

Determinants of pro-environmental behavior have been widely studied in the field of economics and environmental studies. Social norms are an important determinant in behavior; social norms are the collective, perceived, acceptable behavior of the majority. Social norms can be activated through injunctive motivation such as moral obligation. Or norms can be activated descriptive motivation which involves someone’s perception of the status quo.

In the context of recycling, past studies suggest that people’s social norms play a role in determining the likelihood of engaging in recycling programs. Latif and Omar (2017) investigated the recycling behavior correlation to attitude, materialism and collectivism. Using data from a questionnaire, they evaluated relatedness of these factors with a correlation coefficient and found that collectivist attitude was the highest determinant of recycling behavior. Their results imply that people base their recycling decision on what other people are doing.

Other researchers have found correlations between social motivation and recycling behavior. Largo-Wight, Bian, and Lange (2012) examined the determinants of recycling intention on a college campus. They found that the strongest predictor of campus recycling intervention were injunctive norms. Similarly, Izagirre-Olaizola et al (2014) found that altruistic motivation and perceived control had the greatest effect on recycling behavior in a university setting. However both these studies measured descriptive norms based on a questionnaire in which people are asked how much they value others perceptions. However, their measurements are self-reported and therefore may not reflect people’s perception of the social norm nor reflect their recycling behavior.

Other than formal interventions of social motivation, peripheral indicators can also influence someone’s decision to recycle. Abbot, Nandeibam, and O’Shea (2013) examined the role of social norms and warm glow effect on recycling. They measured social awareness by facilitating visibility of recycling efforts in kerbside collection. They found that social norms was the highest predictor of recycling behavior for households. Similarly, Tong et al. (2018) used visibility of household recycling as a measurement of social norm predicting a community’s decision to recycle, and they found that attitude towards recycling had low impact on the likelihood of a community to recycle. The visual representation of the social norm can have an impact on people’s behavior.

Descriptive norms are important determinants of recycling. Using the visual cue to indicate descriptive norms means that a person can be informed about the decisions of others in real-time as they make their decision. This theory of behavior suggests that real-time feedback is more powerful behavior motivation than aggregate simultaneous and doesn’t involve any norm explanation (Biel, 2007). In a meta-analysis of many studies on recycling, Hornik et al (1995) found that perceived social influence has a strong propensity to predict recycling behavior.

III. Model Development

I want to test the effect of social norms on sorting behavior controlling for other related factors. Based on past studies on the subject and descriptive norm theory, I developed the following model to estimate the effect of fullness of the compost bucket on sorting behavior:

Sorting_{i} = β_{0} + β_{1}Fullness_{i} + β_{2}Busyness_{i} + + β_{4}Knowelge_{i} + β_{5}Fullness^{2}_{i} + β_{6}Group + ε_{i}

i= individuals, cross-sectional data

Sorting is a dummy variable 1=sorting and 0= not sorting. There is no partial sorting for people who sort but do it incorrectly. I only tested if people simply put everything in landfill or not, I was not testing people’s knowledge of how to recycle and compost properly. The first determinate I tested was the influence of descriptive norms by measuring the fullness of the compost bucket, β_{1}Fullness_{i}. The assumption is that the more full the bucket is, the more evidence that others are sorting their trash which should indicate to any individual making their decision what the social norm is. While a written sign could tell people that the majority of students recycle, I assumed that most people will not take the time to read a sign when throwing their trash away. People are not actively thinking through their recycling decisions on a regular basis, therefore a quick visual cue of how much people are composting can be a powerful indicator. I would expect the sign of the regression coefficient to be positive because I theorize that the fullness of the bucket has a positive effect on people’s likelihood to sort their trash. I expect increasing returns of sorting behavior with respect to fullness because as the bucket fills, the more evidence there is that this is the norm. Descriptive norm theory suggests that people will do things when they think the majority is doing it as well. However, model (1) tests for linear relationship because theory does not specify the relationship between these variables.

The second determinate I tested was busyness, β_{2}Busyness_{i}. I expect less sorting at times when more people are throwing things away. This assumes that the more frequent people are throwing things away, the busier people are; groups of people tend to leave the union during a class-transition period. I would expect the sign of the regression coefficient for busyness to be negative because the busier people are, the less time and thought they were going to put into their waste decision. I expect the relationship between sorting and busyness to be linear.

I would expect other determinants to also affect waste including knowledge that the individual has on recycling. A college hosts students from across the country with different exposure to recycling. Past research indicates that recycling knowledge is likely an indicator of their sorting behavior (Largo-Wight, Bian, and Lange ,2012) (Biel, 2007). For this reason, I would expect the sign of the regression coefficient to be positive.

I also expect whether the individual is in a group or not to influence their sorting behavior. Theory of Planned behavior suggests that people are more likely to perform an environmentally friendly task when they believe that they are being watched (Biel, 2007). I believe the coefficient of this variable will be positive. However, I would also recognize that people in a group may also negatively impact sorting behavior because they might be talking or distracted when throwing their trash away.

I expect the omission of factors from model (1) described in model (4) to lead to autocorrelation and heteroskedasticity. I expect autocorrelation because I predict that people’s behavior influences others’. For example, if the individual before you sort their trash and you observe this action social influence theory predicts that that will influence your actions. I also accept heteroskedasticity because I expect different groups on campus to have different likelihoods than others. For example, I expect that students with lots of recycling knowledge to have a higher likelihood to sort than others.

Although I expect many factors to influence someone’s likelihood to sort their trash, model (1) will test in this paper will only include two explanatory variables: busyness and fullness. Model (1): Sorting = β_{0} + β_{1}Fullness_{i} + β_{2}Busyness_{i}

Even though I expect fullness to have a non-linear relationship with sorting, model (1) tests for linear relationship because theory does not specify the relationship between these variables.

IV. Discussion of the Data

My target population is Davidson College student body. My sample frame is people eating and dining at the Davis Cafe when I was observing their behavior. This study was conducted over two days. I collected data during lunch time when class periods transitioned. I set out a composting bin at 8 am both days with a sign that says “composting, food scraps only, trial basis.” I and observed people’s behavior from near-by table. I recorded whether they had something to sort and if they sorted. Participants did not know they were a part of the study. I did not need IRB approval because the data does not involve sensitive topics, vulnerable subjects, or personally identifiable information.

Although I recorded the sorting behavior of 100 people who disposed waste, only 84 of them had some compostable material. In my regression analysis, I only used data on people with compostable material because there is no way to test their likelihood to sort if they have nothing to sort. For the total sample, including those without compostable materials, the mean busyness was 10 people within a 10 minute period. Demographic information for everyone who discarded trash are summarized in Table 1 below. People had compostable materials 77.8% of the time and sorted their trash 40.7% of the time. 89.8% of the sample was students and not a faculty or staff. People threw away their trash can in a group 31.5% of the time.

**Table 1: Summary Stats of the Total Sample **

Mean | Standard Error | Median | Mode | Standard Deviation | Sample Variance | Range | Minimum | Maximum | |

business | 10.120 | 0.645 | 8 | 6 | 6.704 | 44.948 | 23 | 1 | 24 |

fullness | 0.236 | 0.024 | 0.25 | 0 | 0.248 | 0.062 | 0.75 | 0 | 0.75 |

composable or not | 0.778 | 0.040 | 1 | 1 | 0.418 | 0.174 | 1 | 0 | 1 |

sorting behavior | 0.407 | 0.048 | 0 | 0 | 0.494 | 0.244 | 1 | 0 | 1 |

student or faculty | 0.898 | 0.029 | 1 | 1 | 0.304 | 0.092 | 1 | 0 | 1 |

in a group? | 0.315 | 0.045 | 0 | 0 | 0.467 | 0.218 | 1 | 0 | 1 |

The sample I performed the regression on consisted only of people who had compostable trash. Table 2 summarizes some statistics only from the part sample I used in the regression model. Of the 84 participants, they sorted their trash 33.33% of the time the mean business was 9.976 which is similar to the total sample, 10.12. Fullness ranges is 75 because I can not record sorting behavior when the bucket is 100 full. The sample variance of fullness is very large 632.172 and business as well 44.746, this indicates that our OLS estimate for causation is strong.

**Table 2: Summary Stats of the Sample **

Mean | Standard Error | Median | Mode | Standard Deviation | Sample Variance | Range | Minimum | Maximum | Confidence Level(95%) | |

Sorting | 0.333 | 0.052 | 0 | 0 | 0.474 | 0.225 | 1 | 0 | 1 | 0.103 |

fullness | 24.405 | 2.743 | 25 | 0 | 25.143 | 632.172 | 75 | 0 | 75 | 5.456 |

business | 9.976 | 0.730 | 7 | 6 | 6.689 | 44.746 | 23 | 1 | 24 | 1.452 |

V. Analyses

I chose a significance level of 5% because I place equal weight on Type I and type II error. Type I error would be if I reject the null hypothesis that fullness has no effect on sorting behavior. While a type II error would be if I found no effect when there was an effect on fullness on sorting behavior. Each of these errors has equal adverse effects, so I chose to weight there significance equally with an alpha of 0.05.

Graph 1 and 2 display the unconditional scatter plots and line of best fit for the two explanatory variables on the variable of interest. There appears to be a positive linear relationship between fullness and sorting in graph 1. Busyness seems to be negatively correlated with sorting. These relationships are confirmed by the correlation coefficients in Table 3.

I also analyzed the correlation between all my variables in model (1) using Pearson’s correlation coefficient. I tested the correlation between Sorting and fullness against the alternative hypothesis that fullness and sorting have a correlation greater than 0 because expect the correlation to be positive. Indeed, sorting and fullness have a weak positive correlation of 0.24 which is significant at the 5% level. I also tested the correlation between sorting and buysness against the alternative that busyness will have a negative correlation coefficient. I found a very weak correlation that is not significant even at the 10% level. Lastly I tests to see if there was any relationship positive or negative between fullness and sorting. I found a weak negative correlation -0.38, but this was significant at the 1% level indicating that there is a relationship between these variables.

**Table 3: Correlation Coefficient Hypothesis Test**

Y and X1 | Y and X2 | X1 and X2 | |

Ho | ρ = 0 | ρ = 0 | ρ = 0 |

H1 | ρ < 0 | ρ > 0 | ρ ≠ 0 |

r | 0.244189** (2.26631) | -0.115204 (-1.04378) | -0.386914*** (-3.77634) |

Table 4 summarizes the results of the OLS regression for model (1) (2) and (3). On average, holding all else constant, as the compost bucket increases by 10 percentage points in fullness leads to a 0.04 percentage point increase in someone’s likelihood to sort their trash. I accepted the sign of the coefficient to be positive which it is. The results are significant at the 5% level. The sample of people with compostable materials about 12% of people sort their trash when the bucket was less than 25% full. These results imply that if I put out a quarter full bucket instead of an empty one every time I put out a compost bin, I can expect to increase the first composter likelihood to sort their trash by almost 50% assuming average busyness.

**Table 4: OLS Regression Results**

Dependent Variable is Sorting Behavior | |||

Independent variable | Model (1) | Model (2) | Model (3) |

Fullness (%) | 0.00443** (0.00220) | – | 0.36006 (0.23722) |

Busyness (Persons) | -0.00173 (0.00828) | -0.00817 (0.00777) | – |

Intercept | 0.24251* (0.12543) | 0.41481*** (0.09323) | 0.22617** (0.08729) |

R² | 0.06013 | 0.01327 | 0.02733 |

Adjusted R² | 0.03693 | 0.00124 | 0.01546 |

Root Mean Square Error | 0.46540 | 0.47394 | 0.47055 |

Number of observations | 84 | 84 | 84 |

***, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively. Standard errors are in parentheses. Fullness is measured where when 0 = the bucket is 0-¼ full, 1= ¼-½ full and so forth, the bucket was removed once full. Business is a dummy variable, business is the number of people who also threw away trash within a ten minute interval when any given individual threw away trash.**

For every ten more people also throwing things away within the ten minute interval of any individual throwing something away, makes a person 0.017 percentage points decrease in probability of sorting behavior. The busyness coefficient is not statistically significant at the 5% level. I accepted a negative sign; I predicted that busyness would have a negative impact on someone’s likelihood to sort. At the peak business times during class period transitions, there were as many as 24 people throwing away their trash in a ten minute period, during these times holding all else constant, we would expect someone’s likelihood to sorting behavior to decrease by 7.3% compared to the average busyness assuming average fullness.

On average when busyness and fullness are 0, then the probability of sorting behavior is 24.5% likely. This is this is significant at the 5% level. I accepted this sign to be positive because at least a few people will sort their trash and it would not make sense to have a negative probability. These results imply that when the bucket is empty and no one has thrown anything away within the ten minute period someone is throwing their trash away, the likelihood to sort is 24.5%.

For the hypothesis tests H_{0}: β_{1 }= 0 (vs. two-tailed alternative), I calculated the power for the alternative true value of β_{1 }= 0.36. I used the effect size from model (3), because the different in coefficients is so different from model (1). The power is 93% at the 5% significance level. This means the model will accept the null hypothesis when it is false only 7% of the time. If the change in fullness has a significant effect on the likelihood someone has to sort their trash, there is a 93% chance that my model will detect it.

Model (1) does not have a good fit to the data. The adjusted R^{2 }is extremely small, only 0.037, indicating a weak relationship between my independent and explanatory variables. Only 3.7% of the variation in sorting can be explained by the fullness and busyness. Also, the estimation of the linear model in the equation is not very accurate. The standard error of the model is 0.46 meaning that on average any one person’s sorting behavior is within 0.46 percentage points of the estimated behavior at the 95% confidence level. The Adjusted R^{2} is small and the standard error is large therefore the model has a low goodness of fit.

There is a difference in the fullness regression coefficient from model (1) and model (3) When busyness is not considered, a 10% increase in fullness leads to the likelihood of someone sorting to increase 3.6 percentage points. This is a 90% increase in the value of the regression when this is omitted. The standard error of this variable also increases drastically. Indicating that busyness is a relevant factor in the estimate of sorting behavior. When Busyness is omitted the standard error of model increases further providing evidence that busyness is related to sorting behavior and should be included in the model. This is not surprising considering I found a significant correlation between fullness and busyness (Table 3).

There are no significant sign changes, providing evidence that there is no multicollinearity. There is no significant changes in the root mean standard error between the models indicating that the correlation between fullness and busyness overlaps with the sorting behavior. The error term does slightly increase when either one of explanatory is omitted, this is expected. Similarly there is a decrease in adjusted R^{2} moving either one of the variables is omitted which indicates that the addition of each variable increases the goodness of fit for the model (1) especially for fullness.

I don’t expect that Gauss-Markov assumption of zero conditional mean holds for model (1). I suspect that there is a correlation between the explanatory variable and the error term because of omitted variable bias. I believe there are other variables that explain sorting behavior such as the ones included in model (4). I also suspect that model (1) has the wrong function form because I believe there may be increasing returns to sorting behavior as the bucket fills.

There is evidence for heteroscedasticity for both fullness and busyness variables in Graph 3 and 4. Variance in residual is not constant at every level of fullness indicating that sample variance of error is an unbiased estimator of population variance, 𝑉𝑎𝑟(𝜀 l 𝑋) ≠ 𝜎_{ε}². The heteroskedasticity of fullness could be caused by the skewness of the fullness variable, fullness is moderately positively skewed (Table 2). I tested hetrodascity with a Breush-Pagan test not assuming normal distribution of the error, the chi-square p-value is 0.0625, therefore I reject the null hypothesis at the 5% level providing evidence of heteroskedasticity for fullness. The consequence of heteroskedasticity is that the error term is incorrect. I suggest correcting for this by weighting the higher fullness values which exhibit a smaller variance.

I also tested for serial correlation to see if people sorting their waste at the same time influence the behavior to sort. Graph 5 displays the residual error as a function of time, there are clusters negative residuals in minutes near each other. I used a Durbin-Watson test of the entire data set to detect for first-order correlation. I found the Durbin-Watson critical values are 1.600 and 1.696 my d-test statistic was 1.752, therefore I fail to reject the null hypothesis at the 5% level. Thus, I found no significant evidence for autocorrelation. However, when I run the test only using data from only the first day which is in temporal order, the critical values are 1.364 and 1.590, and my DW test stat is 1.56 and the results are ambiguous.

The model is also linear. Any unit increase in fullness causes an increase in sorting behavior and that increase is the same for every increase in fullness. Likewise, any change in busyness doesn’t lead to a nonlinear change in sorting behavior. Similarly, when sorting behavior changes, both of the explanatory variables change at a constant slope. All of this evidence verifies the assumption that the model is linear holds. Therefore, I can assume that all the variation in sorting is caused by the explanatory variables and the error term.

Finally, the sample is independent and identically distributed. I have a large sample size, and I believe that my sample represents the larger population. I only collected data at lunch time, I expect the populations that eats lunch at the cafe during the time I collected data to represent what the population is most of the time. Therefore, for the most part, every individual in the population had an equal and random probability of being sampled from. Further I expect the outcome of one individual to have an effect on the other individual sample.

Given that my model violates two of the five Gauss Markov assumptions, I can not say that the regressor of fullness is the best linear unbiased estimator of the effect of fullness on sorting behavior. For this reason, I will add relevant variables group to capture some of the serial effects and add a quadratic to fullness because I expect increasing returns to sorting as fullness increases. The results of this model are below.

**Table 5: OLS Regression Results**

Independent variable | Model (1) | Model (2) | Model (3) | Model (4) |

fullness | 0.00443** (0.00220) | – | 0.36006 (0.23722) | 0.02045*** (0.0060) |

fullness^{2} | – | – | – | -0.00024*** (0.00008) |

busyness | -0.00173 (0.00828) | -0.00817 (0.00777) | – | -0.00158 (0.0080) |

group affect | – | – | – | 0.01528 (0.10952) |

Intercept | 0.24251* (0.12543) | 0.41481*** (0.09323) | 0.22617** (0.08729) | 0.14168 (0.12591) |

R² | 0.06013 | 0.01327 | 0.02733 | 0.15222 |

Adjusted R² | 0.03693 | 0.00124 | 0.01546 | 0.10930 |

Root Mean Square Error | 0.4654 | 0.47394 | 0.47055 | 0.44757 |

Number of observations | 84 | 84 | 84 | 84 |

***, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively. Standard errors are in parentheses. Fullness is measured where when 0 = the bucket is 0-¼ full, 1= ¼-½ full and so forth, the bucket was removed once full. Business is a dummy variable, business is the number of people who also threw away trash within a ten minute interval when any given individual threw away trash.**

The adjusted R squared in Model (4) increased when the additional variables were added to Model (1), these variables added to the goodness of fit of the model. The fullness effect is significant at the 5% level using a partial F test. According to the quadratic, the peak sorting behavior is when the bucket is 42% full. Compared to those not in a group, people who are sorting their trash in a group are 0.015 percentage points more likely to sort their trash. These results are not significant at the 5% level. Groups effects do not have a strong impact on sorting behavior. The standard error is smaller in model (4) but still quite large, there are most likely other variables that influence sorting behavior I have not taken into account.

My findings are consistent with other literature on recycling, and show that descriptive norms have an impact on sorting behaviors. Abbot et al (2013) found that social norms increase recycling by about 0.4-1.1 percentage points. Although they had different measurements for recycling behavior and for social norm, and they were about to control for education and other demographic information. Overall, the sign of the regression coefficient is consistent with other studies.

VI. Conclusion

Overall my results indicate that social norms have an impact on people’s behavior. The results of this study are particularly compelling because they reveals that descriptive norms alone can influence someone’s likelihood to sort. The standard error of the regression is large even in model (4) with additional values indicating that there are still other related factors that influence sorting behavior that were not accounted for. I suggest further research that tests for other influencing factors such as environmental motivation, or other value-based theory.

### References

**Abbott, Andrew & Nandeibam, Shasikanta & O’Shea, Lucy. **2013.** “**Recycling: Social norms and warm-glow revisited.” *Ecological Economics* 90. 10–18. 10.1016/j.ecolecon.2013.02.015.

**Biel, Anders & Thøgersen, John. **2007. “Activation of Social Norms in Social Dilemmas: A Review of the Evidence and Reflections on the Implications for Environmental Behaviour.” *Journal of Economic Psychology.* 28. 93-112. 10.1016/j.joep.2006.03.003.

**Izagirre-Olaizola, Julen & Fernández-Sainz, Ana & Vicente-Molina, Maria**. 2014. “Internal determinants of recycling behaviour by university students: A cross-country comparative analysis.” *International Journal of Consumer Studies*. 10.1111/ijcs.12147.

**Largo-Wight, Erin, Hui Bian, and Lori Lange**. 2012. “An Empirical Test of an Expanded Version of the Theory of Planned Behavior in Predicting Recycling Behavior on Campus.”* American Journal of Health Education* 43 (2) (Mar): 66-73. https://ezproxy.lib.davidson.edu/

**Latif, S. A., & Omar, M. S. 2017.** “Determinants of Recycling Behaviour in Tioman Island.” *Journal of Asian Behavioural Studies*, 2(4), 49-57.

**Tong, Xin & Nikolic, Igor & Dijkhuizen, Bob & Hoven, Maurits & Minderhoud, Melle & Wäckerlin, Niels & Wang, Tao & Tao, Dongyan**. 2018. “Behaviour change in post-consumer recycling: Applying agent-based modelling in social experiment.” *Journal of Cleaner Production*. 187. 10.1016/j.jclepro.2018.03.261.