Learning to Avoid Unintended Consequences

Free download. Book file PDF easily for everyone and every device. You can download and read online Learning to Avoid Unintended Consequences file PDF Book only if you are registered here. And also you can download or read online all Book PDF file that related with Learning to Avoid Unintended Consequences book. Happy reading Learning to Avoid Unintended Consequences Bookeveryone. Download file Free Book PDF Learning to Avoid Unintended Consequences at Complete PDF Library. This Book have some digital formats such us :paperbook, ebook, kindle, epub, fb2 and another formats. Here is The CompletePDF Book Library. It's free to register here to get Book file PDF Learning to Avoid Unintended Consequences Pocket Guide.

Zulu home language speakers have a choice to either enrol for this course or to take another course in their chosen field of study. To address this problem we have used regression adjustment and entropy matching to help achieve an appropriate balance between the two groups. Having achieved this balance, any difference in the overall performance between the two groups can then be attributed directly to the new language policy that has been implemented at the university. More specifically, in this paper, we will be wanting to consider the effect that this new language policy is having on three different cohorts of students; namely.

Enrolment figures for each of these student groups are given in Table 1. For Black African students who speak another African language, not being able to take an extra course in their chosen field of study may impact negatively on the results they could be achieving for their other subjects had the new language rule not been in place. Many African languages however have a common root. Consequently having to complete a course in Zulu may not be as difficult a learning process for them as would be the case for a European or Indian student who would also have to eventually complete a course in Zulu at UKZN.

Because English is the medium of instruction that is being used at UKZN, should the Zulu speaking students not also be asked to take a similar type of bridging course in English to help them better understand some of the concepts that will be taught in the other courses that they will be taking?

Local studies [ 5 — 6 ] have found that a culture of rote learning is widely prevalent amongst many second language learners in South African universities.

Learning to Avoid Unintended Consequences

Asking a non-Zulu home language speaker to complete a course in Zulu has a clearly defined socio-economic benefit for them. Asking a Zulu home language speaker to complete a similar course in English may help them to overcome a significant language barrier that is being hidden behind a culture of rote learning. Let Y i represent a weighted mean mark for all the exams that student i writes in Observations on Y i and T i can then be used to compute the following expression 1 which essentially subtracts a sample mean of outcomes on Y i for students in the non-treated group from a sample mean of outcomes on Y i for students in the treated group.

If the allocation of students to a treated or untreated group is being done in a random manner then 1 provides one with an estimate for an effect that can be attributed directly to the treatment variable alone. Student allocation to a treated or non-treated group however is not being done on a random basis. Whereas Zulu speaking students have the option to enrol for this course, students who have not done Zulu at school have to eventually complete this course as part of their degree structure.

Ideally one would like to be able to observe a weighted mean mark Y 1 i that student i obtains for all their subjects if they took Zulu in and also a weighted man mark Y 0 i had that same student chosen not to take Zulu in The following expression could then provide one with a sample based estimate of the effect on academic performance that the new language policy is having 2. Unfortunately, the above expression can never be calculated because only one of the outcomes Y 1 i and Y 0 i can actually be observed once a student has made a decision T i to take or not to take Zulu in Fortunately, given some additional assumptions and are two causal effects that one may be able to estimate using the data Y i that one is able to observe.

In the context of this paper, ATE represents an average treatment effect on Y i that one can associate with every student who enrols for a degree in UKZN during whereas ATT represents an average treatment effect that applies only to those students who have chosen to take Zulu as part of their curriculum in A negative value for ATT would indicate that had these students in this treated group not been required to take a course in Zulu then they would have done better in their other courses. A treatment assignment mechanism is said to be strongly ignorable if T i and Y 0 i , Y 1 i become independent after conditioning on a set of background variables X.

These results underpin a regression adjustment method that can then be used to estimate ATT and thus determine whether or not the new language policy is adversely affecting the performance of students who are having to complete a course in Zulu as part of their degree. Details of the estimation method are given in an appendix to this paper. Matching is another nonparametric method that one could consider using to adjust for the confounding effect that a set of background variables X may be having on our outcome variable Y. With an appropriate balance having been achieved, controlling further for X becomes unnecessary since it is now unrelated to the treatment variable T and thus a difference in the weighted means on the matched data can be used to estimate the causal effect.

For example, 3 can be used to estimate ATT with N T representing the number of observations in the treated group that have been collected and w ij the weight that is being assigned to a non-treated outcome j based on an appropriate match with an observation i in the treated group. Typically the matching of an observation in the treated group with one or more observations in the control group requires the calculation of a propensity score or an appropriately chosen distance measure [ 11 — 20 ].

One then has to check whether the balance that is being achieved is adequate and perform another rematch or discarding of units if this is not the case.

Mitigating Against AIs Unintended Consequences

Unfortunately, with each iteration, there is no guarantee that an improvement in balance will be achieved. In addition, ATT estimation can often perform poorly [ 17 ] particularly when there is insufficient overlap in the covariate distributions associated with the treated and control groups. To help overcome these problems, Hainmueller [ 20 ] has developed a reweighting scheme for the non-treated observations that asks the user to pre-specify a desired level of covariate balance that they want to achieve between the covariates in the treated and control group.

For example, one may want a balance between the treated and untreated populations on the mean, variance and third moment of a chosen set of background variables X to be achieved.

Subject to these balance conditions being satisfied, a set of weights w ij for the non-treated observations in 3 are then derived that minimize an appropriately chosen Kullback-Leibler entropy score. Because the balancing constraints are being directly built into the reweighting scheme that is being implemented, a check for balance after each reweighting no longer becomes necessary.

The weights that result from this entropy balancing can then be used as inputs for estimating a desired treatment effect. The academic performance of 48, students enrolled for a degree at UKZN in formed the basis for this study. The data came from the University of KwaZulu-Natal and the study has been approved by that universities Ethics committee and Research committee. The year was chosen because that was the year in which the new language policy was first implemented.

An appropriate measure for academic performance was obtained by adding up the percentage mark obtained for each subject taken in and dividing that by the total number of subjects written over the year. Students take a total of eight courses over the year. These courses are split into two half year semesters with the Zulu course being able to be taken in either semester.

If Zulu is taken then it replaces one of the other courses that could be taken as part of their degree structure. A summary of results obtained for the treated and non-treated groups is given in Table 2. A two-sample t-test for a difference in means between the marks being recorded by the treated and non-treated groups respectively produced a t- value of 5. Before one can conclude that the new language policy is actually helping them to perform better in all their other studies one needs to make sure that there are no other background variables that may be masking the true effect that the taking of a course in Zulu may be having on the mean marks that students are recording in the treated and control groups.

Table 3 lists the background variables that we have chosen to include in our analysis. In South Africa, schools have been ranked into quintiles based on the resources that they have at their disposal with the quintile 1 schools being the poorest and the quintile 5 schools the richest. The variable Quint12 that appears in Table 2 refers to a student who has come from a poor schooling background.

10 Fascinating Examples of Unintended Consequences

When a student writes their final school leaving exams the results that they obtain are often summarised in the form of an ordinal point score for each subject. The higher the point score, the better the result that they have achieved for that particular subject. These point scores are then added up producing a Matric Point score which we have called MatPts in Table 3. The academic structure at UKZN has been divided into four colleges.

  • Show and Tell & The Great Ice Cream Caper (Autobiography of a Sixth Grader Book 1).
  • Of Moose and Men: Home is Where the Harm Is;
  • Toys at War (Silly Lilly Series)!
  • The Turnaround;
  • Supplemental Content!

Binary indicator variables will be used to model the effect of these colleges in the results that follow. To avoid identification issues, however, only the indicator variables for the first three colleges will be used with an effect for the College of Law and Management Studies forming part of the intercept term in each model structure. To check for a possible imbalance in the distribution of background variables between the treated and non-treated groups, appropriate proportions and variances were calculated for each of the background variable being given in Table 3.

A test for a difference in proportions between the treated and non treated groups using the following asymptotically normally distributed test statistic was also done. The results in Table 4 indicate a strong imbalance in the covariate distributions between the treated and control groups for all covariates except possibly the covariate Afzulu indicating whether the student is Zulu home language speaker or not.

Estimates for the mean treatment effects that are given in Table 4 indicate that having to take a course in Zulu is actually lowering the overall weighted mean mark that one can expect to record by more than 2 percentage points with this effect being stronger amongst the treated subpopulation who have actually chosen to take Zulu in This contradicts an earlier conclusion that we would have been made see Table 2 had we chosen to make no adjustment for potential confounders in our analysis. Ordinary least squares applied to model structure A3 that appears in the supporting appendix produced the results that are given in Table 5 and Table 6.

The Stata code that was used to generate these results also appears in the supporting information section of this paper. Females appear to perform significantly better than males with the effect being stronger in the treated subpopulation. This outcome may be reinforcing a notion that females are better at learning a new language bearing in mind that this is an effect that is being recorded after an adjustment for all the other included variables in the model has been made.

Not unexpectedly, students with a higher Matric point score perform better than those with a lower point score with the effect being slightly more stronger in the treated subpopulation.

Chris Adamson’s Blog: Avoid the Unintended Consequences of Analytic Models

Interestingly enough, race being a Black African becomes insignificant once an adjustment for being a Zulu home language speaker Afzulu has been made. Within both groups, Zulu home language speakers significantly underperform when compared with those who speak another home language with the effect being stronger in the treated group. Failure to account for base rates. When we neglect to consider how the past will affect the future, we are failing to account for base rates.

Lessons learned from unintended consequences about erasing the stigma of mental illness.

Schieffelin likely failed to consider the base rates of successful species introduction. We sometimes perform actions out of curiosity, without any idea of the potential consequences. The problem is that our curiosity can lead us to behave in reckless, unplanned, or poorly thought-through ways. The tendency to want to do something.

We are all biased towards action. The problem is that sometimes doing nothing is the best route to take. We cannot eliminate unintended consequences, but we can become more aware of them through rational thinking techniques. In this section, we will examine some ways of working with and understanding the unexpected. Note that the examples provided here are simplifications of complex issues.

  1. The Causes of Unintended Consequences?
  2. All About George Washington A Pictorial Biography for Students.
  3. Search This Blog;
  4. Don’t be evil. Avoid unintended consequences. Get the Ethical OS Toolkit..
  5. The observations made about them are those of armchair critics, not those involved in the actual decision making. When we invert our thinking, we consider what we want to avoid, not what we want to cause. Rather than seeking perfection, we should avoid stupidity.

    Top Stories Past 30 Days

    By considering potential unintended consequences, we can then work backward. For example, the implementation of laws which required cyclists to wear helmets at all times led to a rise in fatalities. People who feel safer behave in a more risky manner. If we use inversion, we know we do not want any change in road safety laws to cause more injuries or deaths. So, we could consider creating stricter laws surrounding risky cycling and enforcing penalties for those who fail to follow them.

    Another example is laws which aim to protect endangered animals by preventing new developments on land where rare species live. Imagine that you are a landowner, about to close a lucrative deal. You look out at your land and notice a smattering of endangered wildflowers.