# On Conditional Bias and Correlation: A Guest Post

After I posted about “Observational Systems: Correlations with Value-Added and Bias,” a blog follower, associate professor, and statistician named Laura Ring Kapitula (see also a very influential article she wrote on VAMs here) posted comments on this site that I found of interest, and I thought would also be of interest to blog followers. Hence, I invited her to write a guest post, and she did.

She used R (i.e., a free software environment for statistical computing and graphics) to simulate correlation scatterplots (see Figures below) to illustrate three unique situations: (1) a simulation where there are two indicators (e.g., teacher value-added and observational estimates plotted on the x and y axes) that have a correlation of r = 0.28 (the highest correlation coefficient at issue in the aforementioned post); (2) a simulation exploring the impact of negative bias and a moderate correlation on a group of teachers; and (3) another simulation with two indicators that have a non-linear relationship possibly induced or caused by bias. She designed simulations (2) and (3) to illustrate the plausibility of the situation suggested next (as written into Audrey’s post prior) about potential bias in both value-added and observational estimates:

If there is some bias present in value-added estimates, and some bias present in the observational estimates…perhaps this is why these low correlations are observed. That is, only those teachers teaching classrooms inordinately stacked with students from racial minority, poor, low achieving, etc. groups might yield relatively stronger correlations between their value-added and observational scores given bias, hence, the low correlations observed may be due to bias and bias alone.

Laura continues…

Here, Audrey makes the point that a correlation of r = 0.28 is “weak.” It is, accordingly, useful to see an example of just how “weak” such a correlation is by looking at a scatterplot of data selected from a population where the true correlation is r = 0.28. To make the illustration more meaningful the points are colored based on their quintile scores as per simulated teachers’ value-added divided into the lowest 20%, next 20%, etc. In this figure you can see by looking at the blue “least squares line” that, “on average,” as a simulated teacher’s value-added estimate increases the average of a teacher’s observational estimate increases. However, there is a lot of variability (or scatter points) around the (scatterplot) line. Given this variability, we can make statements about averages, such as “on average” teachers in the top 20% for VAM scores will likely have on average higher observed observational scores; however, there is not nearly enough precision to make any (and certainly not any good) predictions about the observational score from the VAM score for individual teachers. In fact, the linear relationship between teachers’ VAM and observational scores only accounts for about 8% of the variation in VAM score. Note: we get 8% by squaring the aforementioned r = 0.28 correlation (i.e., an R squared). The other 92% of the variance is due to error and other factors.

What this means in practice is that when correlations are this “weak,” it is reasonable to say statements about averages, for example, that “on average” as one variable increases the mean of the other variable increases, but it would not be prudent or wise to make predictions for individuals based on these data. See, for example, that individuals in the top 20% (quintile 5) of VAM have a very large spread in their scores on the observational score, with 95% of the scores in the top quintile being in between the 7th and 98th percentiles for their observational scores. So, here if we observe a VAM for a specific teacher in the top 20%, and we do not know their observational score, we cannot say much more than their observational score is likely to be in the top 90%. Similarly, if we observe a VAM in the bottom 20%, we cannot say much more than their observational score is likely to be somewhere in the bottom 90%. That’s not saying a lot, in terms of precision, but also in terms of practice.

The second scatterplot I ran to test how bias that only impacts a small group of teachers might theoretically impact an overall correlation, as posited by Audrey. Here I simulated a situation where, again, there are two values present in a population of teachers: a teacher’s value-added and a teacher’s observational score. Then I insert a group of teachers (as Audrey described) who represent 20% of a population and teach a disproportionate number of students who come from relatively lower socioeconomic, high racial minority, etc. backgrounds, and I assume this group is measured with negative bias on both indicators and this group has a moderate correlation between indicators of r = 0.50. The other 80% of the population is assumed to be uncorrelated. Note: for this demonstration I assume that this group includes 20% of teachers from the aforementioned population, these teachers I assume to be measured with negative bias (by one standard deviation on average) on both measures, and, again, I set their correlation at r = 0.50 with the other 80% of teachers at a correlation of zero. What you can see is that if there is bias in this correlation that impacts only a certain group on the two instrument indicators; hence, it is possible that this bias can result in an observed correlation overall. In other words, a strong correlation noted in just one group of teachers (i.e., teachers scoring the lowest on their value-added and observational indicators in this case) can be relatively stronger than the “weak” correlation observed on average or overall.

Another, possible situation is that there might be a non-linear relationship between these two measures. In the simulation below, I assume that different quantiles on VAM have a different linear relationship with the observational score. For example, in the plot there is not a constant slope, but teachers who are in the first quintile on VAM I assume to have a correlation of r = 0.50 with observational scores, the second quintile I assume to have a correlation of r = 0.20, and the other quintiles I assume to be uncorrelated. This results in an overall correlation in the simulation of r = 0.24, with a very small p-value (i.e. a very small chance that a correlation of this size would be observed by random chance alone if the true correlation was zero). What this means in practice is that if, in fact, there is a non-linear relationship between teachers’ observational and VAM scores, this can induce a small but statistically significant correlation. As evidenced, teachers in the lowest 20% on the VAM score have differences in the mean observational score depending on the VAM score (a moderate correlation of r = 0.50), but for the other 80%, knowing the VAM score is not informative as there is a very small correlation for the second quintile and no correlation for the upper 60%. So, if quintile cut-off scores are used, teachers can easily be misclassified. In sum, Pearson Correlations (the standard correlation coefficient) measure the overall strength of  linear relationships between X and Y, but if X and Y have a non-linear relationship (like as illustrated in the above), this statistic can be very misleading.

Note also that for all of these simulations very small p-values are observed (e.g., p-values <0.0000001 which, again, mean these correlations are statistically significant or that the probability of observing correlations this large by chance if the true correlation is zero, is nearly 0%). What this illustrates, again, is that correlations (especially correlations this small) are (still) often misleading. While they might be statistically significant, they might mean relatively little in the grand scheme of things (i.e., in terms of practical significance; see also “The Difference Between”Significant’ and ‘Not Significant’ is not Itself Statistically Significant” or posts on Andrew Gelman’s blog for more discussion on these topics if interested).

At the end of the day r = 0.28 is still a “weak” correlation. In addition, it might be “weak,” on average, but much stronger and statistically and practically significant for teachers in the bottom quintiles (e.g., teachers in the bottom 20%, as illustrated in the final figure above) typically teaching the highest needs students. Accordingly, this might be due, at least in part, to bias.

In conclusion, one should always be wary of claims based on “weak” correlations, especially if they are positioned to be stronger than industry standards would classify them (e.g., in the case highlighted in the prior post). Even if a correlation is “statistically significant,” it is possible that the correlation is the result of bias, and that the relationship is so weak that it is not meaningful in practice, especially when the goal is to make high-stakes decisions about individual teachers. Accordingly, when you see correlations this small, keep these scatterplots in mind or generate some of your own (see, for example, here to dive deeper into what these correlations might mean and how significant these correlations might really be).

1. David Berliner on said: