Mysterious cheaters: An exploration of the moral acceptability of cheating on government benefits.
For this project, I used the world values survey to find which attributes predict the marginal effects of saying it’s “ never OK to receive more government benefits than the law assigned to you.” This work was inspired by “Civic Virtue and Labor Market Institutions 2009”.
github link:https://github.com/grantmgeconomics93/values,https://github.com/grantmgeconomics93/values/blob/main/valcode.Rmd
- Introduction
1b. Result Summary
2. explanation of the statistical metrics
3. results
3b my results from extended data and interpretation (MOST IMPORTANT Just READ THIS EVERYTHING ELISE IS FOR THE SAKE OF TRANSPARENCY AND FUTURE EMPLOYERS )
5. Methodology
Introduction
The World Values Survey is done every five years in most countries around the world to measure the economic, religious, cultural, and social attitudes of the people of the world. The survey contains hundreds of questions. This gives social scientists a snapshot of how the world’s people think. In addition, the survey asks for demographic information such as natural identity. This means it is possible to get estimates of people’s values correcting forself-reported characteristics. I used the variables I tried: most common religions, country, age, sex, class employment status, self-perceived economic class, and year the survey was done. To run a probit on the question, “It is never okay to receive more government benefits than the law assigned to you.”I found the biggest differences from the results in the country dummies. This suggests missing factors that explain most of the differences between answers that the variables used did not pick up. This phenomenon was robust even after adding variables like “ natural pride” and “confidence in civil services.” I then added three datasets on corruption, income level with growth rates, and world regions. My results hopefully add humility to the conversation around the relationship between culture and welfare cheating originally explored by “Civic Virtue and Labor Market Institutions 2009”. Summary
Results summary
1,.The regression is described in “Civic Virtue and Labor Market Institutions 2009”.
The first model has a McFadden R² of 0.07571504 this below the 2 recommended for being considered a good model(1). Despite having dozens of regressors, we are left with a model that doesn’t explain much of the variance. The biggest marginal differences were the country dummies. All other regressors were below .1 or 10%I ran the regression with the country dummies, and theMcFadden R² was 0.01903106. This suggests that this model misses something endogenous (check)to countries. I added the variable “confidence in civil services” in the second regression. This addition had surprisingly little effect on the results. Third, I added “ national pride,” which had little impact on the results.
2. When I added the datasets mentioned above (corruption, growth, global region), I could imperfectly approximate the cultural effects.
The original regression of “Civic Virtue and Labor Market Institutions 2009” has a McKelvey & Zavoina’s R² (mz) of 0.14781843 And an adjusted McFadden (amf) of 0.07296777.
The regression with the original variables without the country variable had a (mz) of 0.05692252 And an (amf) 0.02587255.
I then added a region, i.e., EU, South America. This addition resulted in a (mz) of 0.09372513 And an (amf) of 0.04359901.
Next, I wanted to see how corruption and growth affected the results. I took out any geographic information from the regression. This resulted in an (mz) of 0.07531244 and an (amf) of 0.03499408.
I regressed just on region, resulting in an(mz) of 0.05100049 And an (amf) of0.02303353.
I added region to this regression, resulting in a (mz) 0.13778385 And a (amf)0.06728311.
2. explanation of the statistical metrics
Since this project used a probit, I could not use the traditional r². The McFadden method is widely used in probit papers. However, McFadden, “ as Long (1997) observed, there is no clear interpretation of values other than zero and one; in other words, values between these two ranges seem somewhat arbitrary since there is no meaningful way to determine if they are large or small. Moreover, Hagle and Mitchell II (1992) observed that R2 (mf) significantly underestimates R2 (ols) of an underlying continuous model”(1). There is a rule of thumb you can use to interpret McFadden “adden’s R2 (mf)-values between 0.2 and 0.4 are taken to represent a very good fit of the model (McFadden, 1974). Simulations by (Domencich et al., 1975). (1975) equivalence this range to 0.7 to 0.9 for a linear model (Louviere et al., 2000)(1)”. For this project, the regular mcfadden is insufficient for the extended dataset due to the sensitivity of comparing regression with different numbers of categories. Thus, I used McKelvey & Zavoina’s R2 (mz), which “reportedly approximated R2 (ols) very well in previous studies (see, e.g., Hagle and Mitchell II (1992); Windmeijer (1995); Veall and Zimmermann (1992)). (1)” While not a perfect analog tests show “there is a drop in the numbers for all the measures in the binary case, the modified R2 (r=2) and again McKelvey & Zavoina’s R2 (mz) still perform best. (1)” While not relying on a MLE ratio, I decided that the mz would give the most insightful results. There are experimental R² that are reliable and use a MLE ratio; however, I felt that was outside this project’s scope.
Section 1 Findings and variations
I tried three different probits with the original dataset.
- counties dummies below
While the effects are minor, some significant values exist, such as Protestant And Catholicism. Protestant And Catholics have a 6.23% And 2.65% higher probability of putting 1, respectively. However, all other religious beliefs were insignificant predictors. This finding is corroborated by “Does Religious Belief or Attendance Matter More for Economic Values? “(2023). Which found that “Attitudes in favor of patriarchy are prevalent across all religious traditions, and lawfulness appears to concentrate within Protestantism and Catholicism.” I tried adding the variable asking, “Do you believe in hell?”. I wondered if the belief in cosmic punishment would affect attitudes toward cheating on government benefits. I took out the political identity variable to avoid making the model too complex. The new regression only resulted in a minor change in Protestant with a coefficient of 0.0603987. However, Catholic went down to 0.0156575752(results available on GitHub). The McFadden r² is 0.07531735.
The “low socio-economic” self-identifying individual variable has a significantly negative effect. However, it is just -0.0142.
Model 2
I added the variable “ confidence in civil services.” I wanted to see if faith in civil services would affect people’s willingness to take more government services than allotted. Once again, I deleted the political position so that it did not havetoo many variables. Although this variable was highly significant, it was only -0.0129,0.0101, and -0.0166 for mid-confident, highly confident, and low confident, respectively. I got an adjusted McFadden of 0.07212425.
Model 3
I took out the class variable, and I added national pride. The intuition for this variable addition was low national pride might have had an impact on how people look at welfare. If someone believes their country is not worthy of pride, they might be more likely to cheat. However, this had a negligible effect with a compatible adjusted McFadden.
3B.Expanded dataset analysis and interpretation
The model from “Civic Virtue and Labor Market Institutions” had a mz 0.14781843, the highest among all the regressions I tried. However, it’s still relatively low considering the complexity of the model. I wondered if there’s an honesty problem. However, “Hainmueller, Hangartner, and Yamamoto (2015) survey the literature and provide evidence that survey responses to a range of questions correspond surprisingly well with actual behavior”.Although, there is evidence that how the question is asked is important to getting valid results(see lit review). So, it remains a mystery as to what the other factors explain the majority of the variation in the dependent variable. Still, we gain insights. The second regression I ran, with the extended data set, did not have any geographic or economic information and had an mz of 0.05692252. Though an imperfect methodology, it suggests that ~38% of the variation of the original regression can be explained without cultural factors. The third regression I ran added a region variable to the second regression, and the mz rose to 0.09372513, adding the region variable, which captures cultural factors, explaining 63% of the original regression, in a lower resolution than adding a dummy for every country. I then took out the geographic data and put in economic indicators for the growth of GDP per capita in the 1990s, the corruption index in 2000, and two dummies for having a GDP per capita above 15000 in 2000 and 1990, respectively.
With 1990(inclevalnine0) full model mz 0.07846242
Without 1990 mz 0.07531244
Without the income dummy(dummy for GDP per capita above 15000), icleval mz 0.06914581
without the growth per capita in the 1990s growth (growth) mz 0.05752887
Without any economic data, ie. In the second regression plus the corruption(score), I ran mz 0.05692252.
When I regressed County alone along with a time dummy (year) 0.11491894
I wanted to see how much I could explain without cultural factors. There are several weaknesses of this regression. Due to a lack of data access, I couldn’t get the income level in 1980, the first survey period. So, I think this is a low estimate of how much you can explain with corruption and economic factors alone. Corruption mattered surprisingly little. Some might argue that corruption might be correlated with culture. However, in my dataset, the correlation between corruption and income level in 2000 is -0.5693558. The relationship between corruption and income level might have two-way causality. Countries can’t grow due to corruption, and they have corruption because they have a low income. Interestingly, the correlation between the average income level in a country and the proportion of people who said it is never OK to cheat on government welfare is 0.067 66228. This backs up the notation that we only explain a small portion of the variance. When I added regions in the regression, I got 0.13774276, not far from the 0.14781843 mz of the model in “ “Civic Virtue and Labor Market Institutions”. This is most likely due to a higher resolution of cultural differences. The addition of regions improves the model, but it still only explains a small minority of the variation in the outcome variable. While the mz is not perfect, even if it underestimated the explained variation by a factor of two, the cultural factors influencing the attitude towards cheating on government programs would still be about 10% or twice the mz of regressing regions alone 0.05100049.
We see different effects if we compare the regional marginal effects from the corruptionoplusecon and region alone.
Regionalone
Corruptionplusecon
Comparring the two models we can see the regions the marginal effects changed widely. Yet, the corruptionplusecon model only added 8% explainability.. It’s important to emphasize that the exact numbers of mz are not completely reliable we can draw insights from the levels of the models relative to each other.
corruptionoplusecon
Comparing the original regression from “Civic Virtue and Labor Market Institutions” to econpluscorruption we see similar results, suggesting that culture only plays a minor role in the outcome variable. When interpreting results for real-world policy prescriptions, you need to make sure you explain enough of the variation to conclude from. I don’t think any of the models are sufficient to inform policy.
4. methodology
Country: this was the easiest variable to code. I just mapped the number to the country name using the function “case when.” Every consecutive survey had more countries, so despite the year dummy, there might be an implicit bias because higher-income countries have been in the survey longer.
2. Socio-economic status of the respondent. The first survey had three categories: upper class, upper middle, and lower middle. This is problematic because if you identify as lower class, you might not put an answer. It was potentially oversampling to the higher classes. I mapped upper class to upper class, upper middle to middle class, and lower middle to lower class. The second survey was the most subjective to code. It had upper, middle, skilled, and unskilled. I grouped skilled and middle to middle class. Although I don’t think this is perfect, it’s the best. Survey 3 had four categories: upper, upper-middle, working, and lower class. I grouped upper and upper middle to upper, working, and lower to lower. The fourth survey has the same scheme as the third survey.
3. Religion. I mapped the number of the survey to the survey to the number it had in survey one. I also mapped every other religion not in the survey to “other.
4. Employment: There are three categories: inactive, employed, and unemployed. The inactive category included students, housewives (outdated cringe, there can be househusbands), and others. I grouped anyone who claimed to work an hour with employed. This might lead to misleading results because the line between unemployed and severely underemployed is unclear. Lastly, I mapped unemployment to unemployment.
5. Education: The education categories were recorded by the year you finished your education. Earlier surveys had it caped at age 21; however, later, they would allow you to put any age up to 99. I just capped all survey categories at 21. I got estimates close to “Civic Virtue and Labor Market Institutions.”
6. gender binary men and women
7. Political position left vs. Right on a 1:10 spectrum. I put five as a center with 6:10 right and 1:4 left. I would have liked more options like 1:100 to separate centrists because I could have put 45:55 as centrists. Also, the idea of what is left and right is not identical worldwide. So, this can introduce some noise to the data.
8. Corruption as defined by https://www.transparency.org/en/cpi/2022
9. regions as defined by https://unstats.un.org/sdgs/report/2019/regional-groups/
Model 1
Model 2
Model 3
quotes sources
(1)A Modification of McFadden’s R2 for Binary and Ordinal Response Models
(2)Validating vignette and conjoint survey experiments against real-world behavior
Section two
Lit review
Ajzenman, Nicolás. 2021. “The Power of Example: Corruption Spurs Corruption.” American Economic Journal: Applied Economics, 13 (2): 230–57.
Martinangeli, A.F. and Windsteiger, L., 2022. The Propagation of Unethical Behaviours: Cheating Responses to Tax Evasion.
Bjørnskov, C., 2021. Civic honesty and cultures of trust. Journal of Behavioral and Experimental Economics, 92, p.101693.
Hainmueller, J., Hangartner, D. and Yamamoto, T., 2015. Validating vignette and conjoint survey experiments against real-world behavior. Proceedings of the National Academy of Sciences, 112(8), pp.2395–2400.
Data sources