Statistics, slope, cosine, sine, sign, significance and R-squared

The following applies to elections for Parliament, say for the US House of Representatives or the UK House of Commons, and it may also apply for the election of a city council. When the principle is one man, one vote then we would want that the shares of “seats won” would be equal to the shares of “votes received”. When there are differences then we would call this inequality or disproportionality.

Such imbalance is not uncommon. At the US election of November 8 2016, the Republicans got 49.1% of the votes and 55.4% of the seats, while the Democrats got 48% of the votes and 44.6% of the seats. At the UK general election of June 8 2017, the Conservatives got 42.2% of the votes and 48.8% of the seats while Labour got 39.9% of the votes and 40.3% of the seats (the wikipedia data of October 16 2017 are inaccurate).

This article clarifies a new and better way to measure this inequality or disproportionality of votes and seats. The new measure is called Sine-Diagonal Inequality / Disproportionality (SDID) (weblink to main article). The new measure falls under descriptive statistics. Potentially it might be used in any area where one matches shares or proportions, like the proportions of minerals in different samples. SDID is related to statistical concepts like R-squared and the regression slope. This article looks at some history, as Karl Pearson (1857-1936) created the R-Squared and Ronald A. Fisher (1890-1962) in 1915 determined its sample distribution. The new measure would also be relevant for Big Data. William Gosset (1876-1937) a.k.a. “Student” was famously unimpressed by Fisher’s notion of “statistical significance” and now is vindicated by descriptive statistics and Big Data.

The statistical triad

Statistics has the triad of Design, Description and Decision.

  • Design is especially relevant for the experimental sciences, in which plants, lab rats or psychology students are subjected to alternate treatments. Design is informative but less applicable for observational sciences, like macro-economics and national elections when the researcher cannot experiment with nations.
  • Descriptive statistics has measures for the center of location – like mean or median – and measures of dispersion – like range or standard deviation. Important are also the graphical methods like the histogram or the frequency polygon.
  • Statistical decision making involves the formulation of hypotheses and the use of loss functione to evaluate that hypotheses. A hypothesis on the distribution of the population provides an indication for choosing the sample size. A typical example is the definition of decision error (of the first kind) that a hypothesis is true but still rejected. One might accept a decision error in say 5% of the cases, called the level of statistical significance.

Historically, statisticians have been working on all these areas of design, description and decision, but the most difficult was the formulation of decision methods, since this involved both the calculus of reasoning and the more complex mathematics on normal, t, chi-square, and other frequency distributions. In practical work, the divide between the experimental and the non-experimental (observational) sciences appeared insurmountable. The experimental sciences have the advantages of design and decisions based upon samples, and the observational sciences basically rely on descriptive statistics. When the observational sciences do regressions, there is an ephemeral application of statistical significance that invokes the Law of Large Numbers, that all error approximates the normal distribution.

This traditional setup of statistics is being challenged in the last decades by Big Data – see also this discussion by Rand Wilcox in Significance May 2017. When all data are available, and when you actually have the population data, then the idea of using a sample evaporates, and you don’t need to develop hypotheses on the distributions anymore. In that case descriptive statistics becomes the most important aspect of statistics. For statistics as a whole, the emphasis shifts from statistical decision making to decisions on content. While descriptive statistics had been applied mostly to samples, Big Data now causes the additional step how these descriptions relate to decisions on content. In fact, such questions already existed for the observational sciences like for macro-economics and national elections, in which the researcher only had descriptive statistics, and lacked the opportunity to experiment and base decisions upon samples. The disadvantaged areas now provide insights for the earlier advantaged areas of research.

The key insight is to transform the loss function into a descriptive statistic itself. An example is the Richter scale for the magnitude of earthquakes. It is both a descriptive statistic and a factor in the loss function. A nation or regional community has on the one hand the cost of building and construction and on the other hand the risk of losing the entire investments and human lives. In the evaluation of cost and benefit, the descriptive statistic helps to clarify the content of the issue itself. The key issue is no longer a decision within statistical hypothesis testing, but the adequate description of the data so that we arrive at a better cost-benefit analysis.

Existing measures on votes versus seats

Let us return to the election for the House of Representatives (USA) or the House of Commons (UK). The criterion of One man, one vote translates into the criterion that the shares of seats equal the shares of votes. We are comparing two vectors here.

The reason why the shares of seats and votes do not match is because the USA and UK use a particular setup. The setup is called an “electoral system”, but since it does not satisfy the criterion of One man, one vote, it does not really deserve that name. The USA and UK use both (single member) districts and the criterion of Plurality per district, meaning that the district seat is given to the candidate with the most votes – also called “first past the post” (FPTP). This system made some sense in 1800 when the concern was district representation. However, when candidates stand for parties then the argument for district representation loses relevance. The current setup does not qualify for the word “election” though it curiously continues to be called so. It is true that voters mark ballots but that is not enough for a real election. When you pay for something in a shop then this is an essential part of the process, but you also expect to receive what you ordered. In the “electoral systems” in the USA and UK, this economic logic does not apply. Only votes for the winner elect someone but the other votes are obliterated. For such reasons Holland switched to equal / proportional representation in 1917.

For descriptive statistics, the question is how to measure the deviation of the shares of votes and seats. For statistical decision making we might want to test whether the US and UK election outcomes are statistically significantly different from inequality / proportionality. This approach requires not only a proper descriptive measure anyway, but also some assumptions on the distribution of votes which might be rather dubious to start with. For this reason the emphasis falls on descriptive statistics, and the use of a proper measure for inequality / disproportionality (ID).

A measure proposed by, and called after, Loosemore & Hanby in 1971 (LHID) uses the sum of the absolute deviations of the shares (in percentages), divided by 2 to correct for double counting. The LHID for the UK election of 2017 is 10.5 on a scale of 100, which means that 10.5% of the 650 seats (68 seats) in the UK House of Commons are relocated from what would be an equal allocation. When the UK government claims to have a “mandate from the people” then this is only because the UK “election system” is so rigged that many votes have been obliterated. The LHID gives the percentage of relocated seats but is insensitive to how these actually are relocated, say to a larger or smaller party.

The Euclid / Gallagher measure proposed in 1991 (EGID) uses the Euclidean distance, again corrected for double counting. For an election with only two parties EGID = LHID. The EGID has become something like the standard in political science. For the UK 2017 the EGID is 6.8 on a scale of 100, which cannot be interpreted as a percentage of seats like LHID, but which indicates that the 10.5% of relocated seats are not concentrated in the Conservative party only.

Alan Renwick in 2015 tends to see more value in LHID than EGID: “As the fragmentation of the UK party system has increased over recent years, therefore, the standard measure of disproportionality [thus EGID] has, it would appear, increasingly understated the true level of disproportionality.”

The new SDID measure

The new Sine-Diagonal Inequality / Disproportionality (SDID) measure – presented in this paper – looks at the angle between the vectors of the shares of votes and seats.

  • When the vectors overlap, the angle is zero, and then there is perfect equality / proportionality.
  • When the vectors are perpendicular then there is full inequality / disproportionality.
  • While this angle variates from 0 to 90 degrees, it is more useful to transform it into sine and cosine that are in the [0, 1] range.
  • The SDID takes the sine for inequality / disproportionality and the cosine of the angle for equality / proportionality.
  • With Sin[0] = 0 and Cos[0] = 1, we thus get a scale that is 0 for full inequaliy / disproportionality and 1 for full equality / proportionality.

It appears that the sine is more sensitive than either absolute value (LHID) and Euclidean distance (EGID). It is closer to the absolute value for small angles, and closer to the Euclidean distrance for larger angles. See said paper, Figure 1 on page 10. SDID is something like a compromise between LHID and EGID but also better than both.

The role of the diagonal

When we regress the shares of the seats on the shares of the votes without using a constant – i.e. using Regression Through the Origin (RTO) – then this gives a single regression coefficient. When there is equality / proportionality then this regression coefficient is 1. This has the easy interpretation that this is the diagonal in the votes & seats space. This explains the name of SDID: when the regression coefficient generates the diagonal, then the sine is zero, and there is no inequality / disproportionality.

Said paper – see page 38 – recovers a key relationship between on the one hand the sine and on the other hand the Euclidean distance and this regression coefficient. On the diagonal, the sine and Euclidean distance are both zero. Off-diagonal, the sine differs from the Euclidean distance in nonlinear manner by means of a factor given by the regression coefficient. This relationship determines the effect that we indicated above, how SDID compromises between and improves upon LHID and EGID.

Double interpretation as slope and similarity measure

There appears to be a relationship between said regression coefficient and the cosine itself. This allows for a double interpretation as both slope and similarity measure. This weblog text is intended to avoid formulas as much as possible and thus I refer to said paper for the details. Suffice to say here is that, at first, it may seem to be a drawback that such a double interpretation is possible, yet, on closer inspection the relationship makes sense and it is an advantage to be able to switch perspective.

Weber – Fechner sensitivity, factor 10, sign

In human psychology there appears to be a distinction between actual differences and perceived differences. This is called the Weber – Fechner law. When a frog is put into a pan with cool water and slowly boiled to death, it will not jump out. When a frog is put into a pan with hot water it will jump out immediately. People may notice differences between low vote shares and high seat shares, but they may be less sensitive to small differences, while these differences actually can still be quite relevant. For this reason, the SDID uses a sensitivity transform. It uses the square root of the sine.

(PM. A hypothesis is that the USA and UK call their national “balloting events” still “elections”, is that the old system of districts has changed so gradually into the method of obliterating votes that many people did not notice. It is more likely though that that some parties recognised the effect, but have an advantage under the present system, and then do not want to change to equal / proportional representation.)

Subsequently, the sine and its square root have values in the range [0, 1]. In itself this is an advantage, but it comes with leading zeros. We might multiply with 100 but this might cause the confusion as if it would be percentages. The second digit might give a false sense of accuracy. It is more useful to multiply this by 10. This gives values like on a report card. We can compare here to Bart Simpson, who appreciates low values on his report card.

Finally, when we compare, say, votes {49, 51} and seats {51, 49}, then we see a dramatic change of majority, even though there is only a slight inequality / disproportionality. It is useful to have an indicator for this too. It appears that this can be done by using a negative sign when such majority reversal occurs. This method of indicating majority reversals is not so sophisticated yet, and at this stage consists of using the sign of the covariance of the vectors of votes and seats.

In sum: the full formula

This present text avoids formulas but it is useful to give the formula for the new measure of SDID, so that the reader may link up more easily with the paper in which the new measure is actually developed. For the vectors of votes and seats we use the symbols v and s, and the angle between the two vectors give cosine and then sine:

SDID[v, s] = sign 10 √ Sin[v, s]

For the UK 2017, the SDID value is 3.7. For comparison the values of Holland with equal / proportional representation are: LHID 3, EGID 1.7, SDID 2.5. It appears that Holland is not yet as equal / proportional as can be. Holland uses the Jefferson / D’Hondt method, that favours larger parties in the allocation of remainder seats. At elections there are also the wasted vote, when people vote for fringe parties that do not succeed in getting seats. In a truly equal or proportional system, the wasted vote can be respected by leaving seats empty or by having a qualified majority rule.

Cosine and R-squared

Remarkably, Karl Pearson (1857-1936) also used the cosine when he created R-squared, also known as the “coefficient of determination“. Namely:

  • R-squared is the cosine-squared applied to centered data. Such centered data arise when one subtracts the mean value from the original data. For such data it is advisable to use a regression with a constant, which constant captures the mean effect.
  • Above we have been using the original (non-centered) data. Alternatively put, when we do above Regression Through the Origin (RTO) and then look for the proper coefficient of determination, then we get the cosine-squared.

The SDID measure thus provides a “missing link” in statistics between centered and non-centered data, and also provides a new perspective on R-squared itself.

Apparently till now statistics found little use for original (non-centered) data and RTO. A possible explanation is that statistics fairly soon neglected descriptive statistics as less challenging, and focused on statistical decision making. Textbooks prefer the inclusion of a constant in the regression, so that one can test whether it differs from zero with statistical significance. The constant is essentially used as an indicator for possible errors in modeling. The use of RTO or the imposition of a zero constant would block that kind of application. However, this (traditional, academic) focus on statistical decision making apparently caused the neglect of a relevant part of the analysis, that now comes to the surface.

R-squared has relatively little use

R-squared is often mentioned in statistical reports about regressions, but actually it is not much used for other purposes than reporting only. Cosma Shalizi (2015:19) states:

“At this point, you might be wondering just what R-squared is good for — what job it does that isn’t better done by other tools. The only honest answer I can give you is that I have never found a situation where it helped at all. If I could design the regression curriculum from scratch, I would never mention it. Unfortunately, it lives on as a historical relic, so you need to know what it is, and what misunderstandings about it people suffer from.”

At the U. of Virginia Library, Clay Ford summarizes Shalizi’s points on the uselessness of R-squared, with a reference to his lecture notes.

Since the cosine is symmetric, the R-squared is the same for regressing y given x, or x given y. Shalizi (2015, p18) infers from the symmetry: “This in itself should be enough to show that a high R² says nothing about explaining one variable by another.” This is too quick. When theory shows that x is a causal factor for y then it makes little sense to argue that y explains x conversely. Thus, for research the percentage of explained variation can be informative. Obviously it matters how one actually uses this information.

When it is reported that a regression has an R-squared of 70% then this means that 70% of the variation of the explained variable is explained by the model, i.e. by variation in the explanatory variables and the estimated coefficients. In itself such a report does not say much, for it is not clear whether 70% is a little or a lot for the particular explanation. For evaluation we obviously also look at the regression coefficients.

One can always increase R-squared by including other and even nonsensical variables. For a proper use of R-squared, we would use the adjusted R-squared. R-adj finds its use in model specification searches – see Dave Giles 2013. For an increase of R-adj coefficients must have an absolute t-value larger than 1. A proper report would show how R-adj increases by the inclusion of particular variables, e.g. also compared to studies by others on the same topic.  Comparison on other topics obviously would be rather meaningless. Shalizi also rejects R-adj and suggests to work directly with the mean squared error (MSE, also corrected for the degrees of freedom). Since R-squared is the cosine, then the MSE relates to the sine, and these are basically different sides of the same coin, so that this discussion is much a-do about little. For standardised variables (difference from mean, divided by standard deviation), the R-squared is also the coefficient of regression, and then it is relevant for the effect size.

R-squared is a sample statistic. Thus it depends upon the particular sample. A hypothesis is that the population has a ρ-squared. For this reason it is important to distinguish between a regression on fixed data and a regression in which the explanatory variables also have a (normal) distribution (errors in variables). In his 1915 article on the sample distribution of R-squared. R.A Fisher (digital library) assumed the latter. With fixed data, say X, the outcome is conditional on X, so that it is better to write ρ[X], lest one forgets about the situation. See my earlier paper on the sample distribution of R-adj. Dave Giles has a fine discussion about R-squared and adjusted R-squared. A search gives more pages. He confirms the “uselessnes” of R-squared: “My students are often horrified when I tell them, truthfully, that one of the last pieces of information that I look at when evaluating the results of an OLS regression, is the coefficient of determination (R2), or its “adjusted” counterpart. Fortunately, it doesn’t take long to change their perspective!” Such a statement should not be read as the uselessness of cosine or sine in general.

A part of history of statistics that is unknown to me

I am not familiar with the history of statistics, and it is unknown to me what else Pearson, Fisher, Gosset and other founding and early authors wrote about the application of the cosine or sine. The choice to apply the cosine to centered data to create R-squared is deliberate, and Pearson would have been aware that it might also be applied to original (non-centered) data. It is also likely that he would not have the full perspective above, because then it would have been in the statistical textbooks already. It would be interesting to know what the considerations at time were. Quite likely the theoretical focus was on statistical decision making rather than on description, yet this for me unknown history would put matters more into perspective.

Statistical significance

Part of the history is that R.A. Fisher with his attention for mathematics emphasized precision while W.S. Gosset with his attention to practical application emphasized the effect size of the coefficients found by regression. Somehow, statistical significance in terms of precision became more important than content significance, and empirical research has rather followed Fisher than the practical relevance of Gosset. This history and its meaning is discussed by Stephen Ziliak & Deirdre McCloskey 2007, see also this discussion by Andrew Gelman. As said, for standardised variables, the regression coefficient is the R-squared, and this is best understood with attention for the effect size. For some applications a low R-squared would still be relevant for the particular field.

Conclusion

The new measure SDID provides a better description of the inequality or disproportionality of votes and seats compared to existing measures. The new measure has been tailored to votes and seats, by means of greater sensitivity to small inequalities, and because a small change in inequality may have a crucial impact on the (political) majority. For different fields, one could taylor measures in similar manner.

That the cosine could be used as a measure of similarity has been well-known in the statistics literature since the start, when Pearson used the cosine for centered data to create R-square. For the use of the sine I have not found direct applications, but its use is straightforward when we look at the opposite of similarity.

The proposed measure provides an enlightening bridge between descriptive statistics and statistical decision making. This comes with a better understanding of what kind of information the cosine or R-squared provides, in relation to regressions with and without a constant. Statistics textbooks would do well by providing their students with this new topic for both theory and practical application.

Advertisements

Comments are closed.