A Ratio Test of Interrater Agreement With High Specificity

Abstract

Existing tests of interrater agreements have high statistical power; however, they lack specificity. If the ratings of the two raters do not show agreement but are not random, the current tests, some of which are based on Cohen’s kappa, will often reject the null hypothesis, leading to the wrong conclusion that agreement is present. A new test of interrater agreement, applicable to nominal or ordinal categories, is presented. The test statistic can be expressed as a ratio (labeled Q_A, ranging from 0 to infinity) or as a proportion (labeled P_A, ranging from 0 to 1). This test weighs information supporting agreement with information supporting disagreement. This new test’s effectiveness (power and specificity) is compared with five other tests of interrater agreement in a series of Monte Carlo simulations. The new test, although slightly less powerful than the other tests reviewed, is the only one sensitive to agreement only. We also introduce confidence intervals on the proportion of agreement.

Keywords

interrater agreement test kappa

Introduction

Quantifying interrater agreement is useful in contexts where two raters must judge into what categories a series of observations shall be classified. If agreement between raters is perfect, all judgments will be identical, whereas if both raters use completely different criteria, agreement will occur only by chance. Agreement is often assessed using Cohen’s kappa (Cohen, 1960), for which a value greater than 0.40 is commonly considered a moderate agreement (Landis & Koch, 1977). Recently, Kraemer, Periyakoil, and Noda (2002) explained that a Cohen’s kappa for more than two categories is equal to a weighted average of individual kappas. The individual kappa $κ_{j}$ indicates the agreement of the two raters with respect to category j only. Hence, if one category shows a very strong agreement but the others do not, the mean kappa may nevertheless be high.

As an example, consider the following situation in which two psychiatrists examine N = 100 patients suffering from depression in order to appraise the category of depression (this example is inspired from von Eye, Schauerhuber, & Mair, 2006). The psychiatrists used k = 3 categories of severity. Raters’ categorizations are summarized in Table 1 using a $k \times k$ judgment matrix in which cell ${i, j}$ contains the number of observations that were classified as instances of the ith category by the first rater and of the jth category by the second one. Perfect agreement would result in the main diagonal summing to N, and off-diagonal cells containing zero observations.

Table 1.

Example Data of Ratings Performed by Two Raters (Psychiatrist 1 and Psychiatrist 2) Appraising the Severity of Depression of 100 Patients (N = 100) Using Three Categories of Depression (k = 3).

		Psychiatrist 2
		Category of depression
		1	2	3	Row sums
Psychiatrist 1	1	81	1	1	83
Category of depression	2	1	3	5	9
	3	1	5	2	8
Column sums		83	9	8	N = 100

As seen, both raters share the belief that most patients belong to the first category of severity. This high prevalence does not explain the large number of concordant judgments in that category, as chance would only predict approximately 70 agreement ratings for the first category. Hence, there is good agreement regarding Category 1. On the other hand, the two psychiatrists rarely agree on Categories 2 and 3. For example, Psychiatrist 2 found a total of 8 cases belonging to the third category of severity; of those, only 2 cases are also put in the third category by Psychiatrist 1, a figure below what chance would predict. Instead, the results seem to indicate that what looks like a Level 2 depression to Rater 1 is a Level 3 depression to Rater 2 and vice versa. This suggests disagreement between the two raters on these two categories.

Overall, 86% of the cases are judged identically by the two raters (the cases found in the main diagonal). Should we conclude that, overall, the two raters are in good agreement? Cohen’s kappa suggests a moderate but significant agreement ( $κ = 0.528,$ z = 3.43, p < .001), despite the fact that agreement is missing for two out of three categories.

Similarly, consider the results of Table 2. In this example, 127 cases were classified by two raters in one of 5 categories. Again, agreement is not clear. The raters seem to agree well regarding Categories 2 and 5. They also agree very well on Category 1 (nearly half of the ratings in this category are agreements). However, they agree little regarding Categories 3 and 4. Inspection of the data in Table 2, outside the diagonal, shows that the two raters have opposite interpretations regarding Categories 3 and 4: cases that are instances of Category 3 for Rater 1 are instances of Category 4 for the other. The two raters are not responding randomly, but they are not agreeing. Hence, whereas the rate of agreement is moderately good (32.3%), it does not mean that the two raters agree a little on all categories. The results instead suggest that the raters agree well on a few categories. Overall agreement is missing. Yet Cohen’s kappa is weak but still significantly different from zero ( $κ = 0.148,$ z = 3.30, p < .001).

Table 2.

An Example Where Agreement for Three Categories Accompanies Disagreement for Two Categories.

		Rater 2
		1	2	3	4	5	Row sums ${o_{i,}}_{*}$
Rater 1	1	8	2	1	2	4	17
	2	4	11	5	5	2	27
	3	2	1	5	12	7	27
	4	1	4	15	7	3	30
	5	4	6	2	4	10	26
Column sums	$o_{*, j}$	19	24	28	30	26	N = 127

Note. Contrary to Table 1, there is no significant difference in prevalence between the categories.

These two examples show the need for a statistic of agreement that can distinguish situations with much agreement for a few categories from situations with some agreement for most categories.

The purpose of this article is to present two new measures of interrater agreement, which we call Q _A and P_A . The first is more convenient as it uses the Fisher F tables for critical values, whereas the second is possibly more intuitive, as it is a proportion between 0 and 1. In the next section, we describe these measures, which can be used on nominal or ordinal classifications. Next, we examine the reliability of these measures along with their confidence intervals. Finally, we assess the statistical power but, more important, the specificity of this approach, that is, the ability of a test based on Q _A (or P_A ) to make the difference between mixtures of agreeing and disagreeing category ratings from solely agreeing ratings. The approach developed is akin to analysis of variance, as it partitions variances found in each cell (not just those in the main diagonal) as supporting agreement or supporting disagreement.

In the remainder of this article, the judgments are structured in the form of a square k×k table of observed frequencies. Observed frequencies in cell {i, j} will be noted o_i,j . As per the $χ^{2}$ test on contingency table, we note e_i,j the expected frequency in cell {i, j}, estimated from the marginal frequencies with $e_{i, j} = o_{i, *} \times o_{*, j} / N,$ where $o_{i, *}$ is the total of the ith column and $o_{*, j}$ is the total on the jth row.

Pearson’s chi-square test of independence has been used on occasion to ascertain the significance of the agreement. However, this test only examines the null hypothesis ( $H_{0}$ ) of by-chance co-occurrences in any ${i, j}$ cell. It can be seen as an omnibus test in the sense that any type of nonrandom structure can lead to the rejection of the null hypothesis. Therefore, as already noted by Cohen (1960), a significant χ² statistic calculated from a judgment matrix indicates the presence of any type of category coupling between the two raters, whether they be in agreement (in the main diagonal) or disagreeing (anywhere else in the matrix).

The reader can find a brief description of five alternative tests of interrater agreements in Appendix A. These tests will be compared to the present approach in the subsequent section.

Measures of the Global Structure of the Judgment Matrix

The new approach is a nonparametric one based on the same assumptions as Cohen’s kappa (Brennan & Prediger, 1981). As will be seen, it is sensitive because it uses information outside the main diagonal as well as in the main diagonal. The crucial observation is that high agreement rates in the diagonal cells necessarily imply a shortage of cases outside the diagonal, and conversely. Hence, the whole matrix, not just the diagonal, is informative as to whether there is agreement or not.

The present approach partitions the matrix based on whether the observed cell frequencies deviates from chance (judged by whether the observed count is different from its expected value) and whether such ratings are supportive of agreement or not.

More formally, under $H_{0}$ (no agreement), the observed frequencies o_i ,_j in the k² cells of the judgment matrix should fluctuate near their expected values e_i,_j. Hence, about half of the cells should have a frequency count above their respective e_i,_j, and the other half, below.¹ As in the $χ^{2}$ test, the difference between observed and expected frequencies is standardized with $z_{i, j} = (o_{i, j} - e_{i, j}) / \sqrt{e_{i, j}},$ which under $H_{0}$ and for reasonably large counts is normally distributed.

The first statistic, Q _A, is a ratio of cells favorable to agreement against cells unfavorable to agreement, the former being those that (i) on the main diagonal have counts higher than expected and (ii) off the main diagonal have counts smaller than expected, whereas the latter are those that (iii) on the main diagonal have counts smaller than expected and (iv) off the main diagonal have counts larger than expected. Hence, Q _A is computed as

\begin{matrix} Q_{A} = \frac{Sum of z^{2} over cells supporting agreement}{Sum of z^{2} over cells supporting disagreement} = \frac{(i) + (ii)}{(iii) + (iv)} = \frac{\sum_{i = 1}^{k} {(z_{i, i}^{+})}^{2} + \sum_{i = 1}^{k} \sum_{j = 1, j \neq i}^{k} {(z_{i, j}^{-})}^{2}}{\sum_{i = 1}^{k} {(z_{i, i}^{-})}^{2} + \sum_{i = 1}^{k} \sum_{j = 1, j \neq i}^{k} {(z_{i, j}^{+})}^{2}} \end{matrix}

where the terms in the sums (the components) will be added depending on the sign of the difference between the observed frequency and the expected frequency using

z_{i, j}^{s} = {\begin{matrix} (o_{i, j} - e_{i, j}) / \sqrt{e_{i, j}} & if Sign (o_{i, j} - e_{i, j}) = s \\ 0 & otherwise \end{matrix}

Q_A values range from 0 to +∞, and under $H_{0}$ its central value is about 1. On average under, $H_{0}$ the numerator should contain entries from half of the matrix and is therefore on average a sum of $k^{2} / 2$ squared z scores. Its distribution is thus close to the $χ^{2}$ distribution with ${(k - 1)}^{2} / 2$ degrees of freedom (df). The same is true for the denominator. As a first approximation, the distribution of Q_A is therefore a Fisher F ratio distribution with both df equal to ${(k - 1)}^{2} / 2 .$ Hence, critical values for Q_A can be read from the right tail of the Fisher F ratio distribution.

The second statistic, P_A , is the proportion of variance supporting agreement onto total variance. It can be obtained from the z scores or from Q_A with

\begin{matrix} P_{A} = \frac{(i) + (ii)}{(i) + (ii) + (iii) + (iv)} \\ = \frac{Q_{A}}{1 + Q_{A}} \end{matrix}

The statistic P_A is akin to the eta squared (η²) statistic and ranges from 0 to 1. When the ratings are random, its central value is ½, that is, 50% of the variance in the ratings suggest agreement and the other 50% of the variance suggest disagreement. Hence, a value of ½ supports neither interpretations and indicates random ratings. As found in the literature (e.g., Forbes, Evans, Hastings, & Peacock, 2010), a ratio of the form $X / (1 + X),$ where X follows a Fisher F ratio distribution has a standard beta distribution with parameters a and b dependent on the df of the F ratio distribution, here $a = b = {(k - 1)}^{2} / 4 .$

P_A and Q_A are totally interchangeable. To detect whether the ratings deviate from random, one option is to use Q_A . The null hypothesis is $H_{0}$ : Q_A = 1, and the decision rule is to reject $H_{0}$ if the observed Q_A exceeds a critical value read on an F table with ${(k - 1)}^{2} / 2,$ ${(k - 1)}^{2} / 2$ degrees of freedom. If P_A is used, the null hypothesis is $H_{0}$ : P_A = ½ and the decision rule is to reject $H_{0}$ if the observed P_A exceeds a critical value read from a Beta distribution with parameters a and b equal to ${(k - 1)}^{2} / 4 .$ Note that, in principle, these tests could be left tailed, testing for significant disagreement between the raters, but results to be shown will indicate that they should not be used for that purpose. In between significant agreement and significant disagreement, there is a zone in which the ratings are inconclusive either way, as is the case for random ratings.

Under the null hypothesis, approximately half of the cells, $k^{2} / 2,$ should contribute to the numerator of these statistics and the other half to the denominator. If P_A is larger than ½ (Q_A larger than 1), it suggests that more than one half of the cells contribute to the numerator and less than one half to the denominator. Hence, the distribution of P_A for an observed result different from ½ has Beta distribution with degrees of freedom changed to $a = P_{A} {(k - 1)}^{2} / 2$ and $b = (1 - P_{A}) {(k - 1)}^{2} / 2 .$ From this observation, confidence intervals can be obtained with

C I_{1 - α} of P_{A} = [B_{a, b} (α / 2), B_{a, b} (1 - α / 2)]

where $B_{a, b}$ denotes the quantile function of the Beta distribution with its parameters. The following section will verify this assertion.

Appendix B shows how Q_A and P_A can be computed with the SPSS statistical package. It also shows how to get the p value of the test and confidence intervals for P_A for any level $1 - α .$ Table 3 give some indications on how to interpret P_A .

Table 3.

Interpretation of the Statistics P_A.

P _A	Agreement is . . .
0.40 . . . 0.60	weak or absent
0.60 . . . 0.70	fair
0.70 . . . 0.80	moderate
0.80 . . . 0.90	strong
>0.90	outstanding

An Illustration With Computation of Formulas

We illustrate the computation of the Q_A statistic with an example involving two raters having to examine and codify 200 observations into a system of k = 5 categories, labelled 1 to 5. The judgment matrix and the marginal sums $o_{i, *}$ and $o_{*, j}$ are shown in Table 4, top part; the second part of Table 4 presents the expected theoretical frequencies $e_{i, j} = o_{i, *} \times o_{*, j} / N;$ finally, the standardized differences $z_{i, j} = (o_{i, j} - e_{i, j}) / \sqrt{e_{i, j}}$ are shown in the third part of Table 4.

Table 4.

Cross-Classification Data From Two Simulated Raters, Indicating (Top) the Number of Ratings o_i ,_j in Each i, j Cell, (Middle) Its Corresponding Theoretical Value $e_{i, j}$ , and (Bottom) the Standardized Differences $z_{i, j}$ (N = 200).

Observed frequencies o_i,j
		Rater 2
		1	2	3	4	5	Row sums o_i,*
Rater 1	1	7	5	2	1	3	18
	2	5	13	10	7	8	43
	3	11	4	15	6	9	45
	4	8	11	7	9	6	41
	5	11	5	15	6	16	53
Column sums	o_j*	42	38	49	29	42	N = 200
Expected frequencies e_i,j
		Rater 2					Row sums
		1	2	3	4	5	e_i,*
Rater 1	1	3.78	3.42	4.41	2.61	3.78	18
	2	9.03	8.17	10.53	6.24	9.03	43
	3	9.45	8.55	11.02	6.53	9.45	45
	4	8.61	7.79	10.05	5.95	8.61	41
	5	11.13	10.07	12.98	7.68	11.13	53
Column sums e_*,j		42	38	49	29	42	N = 200
Standardized differences z_i,j
		Rater 2
		1	2	3	4	5
Rater 1	1	+1.656	+0.854	−1.148	−0.997	−0.401
	2	−1.341	+1.690	−0.165	+0.306	−0.343
	3	+0.504	−1.556	+1.197	−0.206	−0.146
	4	−0.208	+1.150	−0.961	+1.253	−0.889
	5	−0.039	−1.598	+0.559	−0.608	+1.460

In this example, the component (i), $\sum_{i = 1}^{k} {(z_{i, i}^{+})}^{2},$ includes all 5 diagonal cells, every cell being positive, and is $1 . 656^{2} + 1 . 690^{2} + 1 . 197^{2} + 1 . 253^{2} + 1 . 460^{2}$ $\approx 10.73$ . Its counterpart, (iii), $\sum_{i = 1}^{k} {(z_{i, i}^{-})}^{2},$ is therefore zero. For the off-diagonal cells, the sum of negative-signed components (ii) is $\sum_{i = 1}^{k} \sum_{j \neq i} {(z_{i, j}^{-})}^{2} = {(- 1.148)}^{2} + {(- 0.997)}^{2} + \dots$ $+ {(- 0.608)}^{2} \approx 11.58,$ and the sum of positive-signed components (iv) is $\sum_{i = 1}^{k} \sum_{j \neq i} {(z_{i, j}^{+})}^{2}$ = ${(0.854)}^{2} + {(0.306)}^{2} + \dots + {(0.559)}^{2} \approx 2.71 .$ Q _A is therefore

Q_{A} = \frac{10.73 + 11.58}{0 + 2.71} \approx 8.23

This value, compared to a F critical value at α = .05, F(8, 8) = 3.438, suggests that the raters are in agreement, Q _A(8, 8) = 8.23, p < .05.

Note that the sum of all the components of Q _A, $10.73 + 11.58 + 0.00 + 2.71$ , equals the usual chi-square test, 25.03. With its ${(k - 1)}^{2} = 16$ degrees of freedom, the critical value being 26.30, we would infer that the data pattern observed in Table 4 does not exhibit any structure and may be ascribed to a pure chance mechanism, $χ^{2} (16)$ = 26.30, p > .05. The chi-square test does not detect agreement in Table 4 because, not being specifically suited for agreement detection, it lacks power and specificity (as our Monte Carlo experiments will confirm).

The second statistic, P_A , is obtained with $8.23 / (1 + 8.23) = 0.892$ , well above the reference value of ½, suggesting a strong agreement. To get confidence intervals, computer software must be used as critical values for the Beta distribution are not given in statistics textbooks. We obtain a 95% confidence interval of [0.62, 0.99], not including ½.

For this example, five statistical tests (the four from Appendix A and Q_A ) make the same decision, rejecting the null hypothesis even at the 0.01 level. However, the same does not happen for the previous examples of Tables 1 and 2 where Q _A alone concludes that agreement is absent. Table 5 summarizes the results for the examples of Tables 1, 2, and 4.

Table 5.

Results of the Five Tests of Agreements and the $χ^{2}$ Test on the Examples of Tables 1, 2, and 4.

	Table 1		Table 2		Table 4
	Test result	p	Test result	p	Test result	p
r	0.860		0.323		0.300
k	0.528		0.148		0.125
Tests on kappa
z _k1	3.43	<.01	3.30	<.01	3.54	<.001
z _k2 = z _k3	6.86	<.01	3.31	<.01	3.58	<.001
Tests on the diagonal elements
z _Σ1 *	9.12	<.001	3.10	<.02	3.16	<.001
z _Σ2	3.23	<.002	3.42	<.01	3.25	<.001
Tests on the structure of the matrix
χ²	8.20	n.s.	57.6	<0.001	25.0	n.s.
Q _A	0.62	n.s.	2.50	n.s.	8.23	<.01
Exact p value	.616		.108		.004
P _A	0.38		0.714		0.892
95% CI of P_A	[.01, .94]		[0.38, 0.95]		[0.62, 0.99]

Note. *Denotes the test assuming the indifference principle. n.s. = nonsignificant at the .05 level.

Reliability of the Statistics and Their Distributions

To examine the merit of the present approach, we explored the statistic P_A using Monte Carlo simulations (the same results were obtained exploring Q_A ). In particular, we examined if the scores’ distributions correspond to their theoretical counterparts. To do so, we generated agreement matrices with various amount of true agreement and checked whether the theoretical 95% confidence intervals contained the results of 95% of the simulations (using the same methodology as in Harding, Tremblay, & Cousineau, 2014). We also checked 99% and 99.9% confidence intervals but the results were comparable and so we do not report the findings.

We manipulated the sample sizes (N, from 50 to 500 by increment of 50), the number of categories k (from 4 to 25), and the true probability of an agreement, ρ_A . Details are given in Appendix C.

A subset of the results is shown in Figure 1. The first thing to note is that the theoretical confidence intervals (shown using error bars) match very closely the limits in which 95% of the simulated P_A rest (shown using gray areas). The theoretical error bars are the least accurate for small number of categories and strong effect size where they overestimate the spread of the results.

Figure 1.

Mean P_A (thick line), 95% confidence intervals (error bars), and spread where 95% of the simulated P_A fell (gray area) as a function of sample size N, when ρ_A is varied from 0 to 0.10 (columns) and k is varied from 5 to 10 (rows). We see that P_A is biased downward in the first column: with no effect, mean P_A should be equal to zero for all N and all k.

Less visible but more critical, the average P_A (as well as the average Q_A , not shown) is biased downward, underestimating the strength of the agreement. This bias is visible in the left panels where there is no agreement (ρ_A = 0), more so for large numbers of categories (e.g., k > 8) and for small sample sizes. We could not find a simple way to undo the bias, which would be suitable to all k. Because of the downward bias, these measures are biased toward detecting disagreement and against detecting agreement. This is why these measures should not be used for detecting disagreement (performing a left tail test of Q_A or P_A ). The bias results in a conservative test (i.e., less powerful than it could be). However, as we will see next, this has little impact on its specificity.

Sensitivity and Specificity of the Test

In order to evaluate the relative merits of the present test of agreement, we ran three more series of Monte Carlo experiments, comparing our ratio test to four other tests of agreement. In the first series, we explored statistical power by manipulating the true rate of agreement ρ_A from 0 to 1. In all the simulations, we manipulated the number of categories k (from 5 to 25), the total sample sizes N (from 50 to 500), and the significance levels (α, .05 and .01). The statistical tests to be compared (described in Appendix A) are the following:

$z_{κ 1},$ Cohen’s simple z test

$z_{κ 2},$ a z test using Fleiss, Cohen, and Everitt’s (1969) more accurate variance approximation

$z_{Σ 2},$ a sum-of-z test not assuming the indifference principle

Q_A , the ratio of agreement test (equivalent to P_A )

$χ^{2},$ the standard χ² omnibus test of independence

We included the $χ^{2}$ test even though it is not a test of agreement to provide a reference. We do not report results for $z_{κ 3}$ (described in Appendix A) as its results are undistinguishable from $z_{κ 2}$ nor do we report results from $z_{Σ 1}$ (for reasons explained in Appendix C).

Apart from being powerful, a good test should also be mostly, if not exclusively, sensitive to agreeing judgments, and not to any other type of consistent categorizations. For example, if observations classified as instances of category 2 by the first rater are systematically put in Category 3 by the second rater (as seen in Table 1), this is a consistent pairing, but not an agreement, and the tests should not reject $H_{0}$ based on situations of this type. The two raters, at least for these two categories, are consistent, but they nevertheless disagree. Specificity is that property of a statistical test in virtue of which it is most sensitive to data configurations relevant to its intended rejection condition but keeps relatively insensitive to other configurations. A test having high specificity should not reject $H_{0}$ at a rate higher than the significance level α when the hypothesized condition of rejection is not present.

The purpose of the last two series of Monte Carlo experiments is to evaluate the specificity of the tests. As before, we manipulated the parameters ρ_C, k, and N. However, in the present context, the parameter ρ_C represents the rate of consistent categorizations between raters. This rate is large when raters have consistent decision rules even if these rules are not in agreement. Overall, with large ρ_C , cases will be more frequent in cells not necessarily on the main diagonal.

In these two series of simulations, there is no overall agreement, but possibly accidental agreements for one or a few categories. Hence, the probability of rejecting $H_{0}$ should be small, otherwise the test is making too many Type I errors. Ideally, the probability of Type I errors should be equal to the significance level (5% in the subsequent figures) and fairly constant across rates of consistent categorizations.

To control precisely for the presence of accidental agreements, the two series present two conditions. In the Random with possible coincident pairings condition, Category “x” from Rater 1 is coupled to some Category “y” of Rater 2, where y may be any category among 1 to k, including by chance the same as Category x (an agreement for this category). In the Random excluding coincident pairings condition, there cannot be a single agreeing pair in the main diagonal.

A detailed description of the simulations’ parameters and algorithms is given in Appendix C.

Results

Because of the very large number of conditions explored, we only report illustrative results, shown in Figures 2 to 4, all using a significance level of .05; these results are typical of what was found in the other conditions and with a significance level of .01. Figure 2 presents the results for one set of parameter (k = 5 and N = 125), one panel per condition. In Figure 3, everything is the same as in Figure 2, except that the sample size is doubled (N =250); in Figure 4, everything is the same as in Figure 2, except that the number of categories is doubled (k = 10).

Figure 2.

(Top panel) Power curve for 5 tests of interrater agreement as a function of the true agreement rate (ρ_A ) when k = 5 and N = 125; (Middle and bottom panels) specificity for the same tests in the same condition as a function of the rate of consistent categorizations (ρ_C ).

Figure 3.

Power curve and specificity curves when k = 5 and N = 250 in the same format as Figure 2.

Figure 4.

Power curve and specificity curves when k = 10 and N = 125 in the same format as Figure 2.

As seen, the results follow three patterns.

Chi-Square Test

As expected, all our results disqualify the $χ^{2}$ test as a test of agreement. First, this test is the least powerful. In Figure 2, top panel, for example, the four other tests reach a power of 50% at ρ_A ≈ 0.07 to 0.09, while chi-square needs to wait until ρ_A ≈ 0.15 to attain it. Second, the $χ^{2}$ test is the least specific, rejecting $H_{0}$ even when there is not a single coincident pairing (Random excluding coincident pairings condition).

Cohen’s ( $z_{κ 1}$ ), Fleiss’ ( $z_{κ 2}$ ), and the Sum-of-z ( $z_{Σ 2}$ ) Tests

For the next three tests, Cohen’s simple formula ( $z_{κ 1}$ ), Fleiss et al.’s ( $z_{κ 2}$ ), and $z_{Σ 2}$ behave roughly in the same manner, be it for the power (top panels of Figures 2 to 4) or the specificity results (middle and lower panels of Figures 2 to 4). It is interesting to note that Fleiss et al.’s intricate formula for the variance of $κ$ , $va r_{2} (κ),$ despite its good performances, does not seem worth the effort as compared to Cohen’s much simpler one. The three tests display the best performance on power, surpassing the new Q_A test, more so when the number of categories is large. However, they are also markedly worse than the Q_A test regarding specificity (sometimes by as much as 40%). Hence, when the raters make disagreeing judgments in a consistent manner, those tests will spuriously signal agreement, an unwanted behavior. Indeed, under the condition of Random with possible coincident pairings (middle panels of Figures 2, 3, and 4), $z_{κ 1}$ (like its variant $z_{κ 2}$ ) and $z_{Σ 2}$ react positively to ρ_C , the rate of consistent categorizations, eating away its specificity. Only in conditions where not a single coincident pair is allowed (bottom panels of Figures 2, 3, and 4) do the three tests perform almost identically to the Q_A test.

Q_A Test

As can be seen in the top panels of Figures 2 and 3, power curves of the Q_A test follow well the power curves of the $z_{κ 1}$ , $z_{κ 2}$ , and $z_{Σ 2}$ tests more so if the number of categories is smaller. In the worst case presented (k = 10, top panel of Figure 4), a power of 50% is attained by Q_A at ρ≈ 0.046, and by Cohen’s test at ρ≈ 0.026. This loss of power continues to increase as k increases. On the other hand, the specificity curves present a remarkable stability for Q_A : the significance rate in all cases starts at the alpha value (e.g., .05) or below and declines steadily as the rate of consistent categorizations increases.

The Q_A test considers all circumstances in which the raters may agree and disagree. Hence, it rejects the null hypothesis only if the evidence gathered from the agreeing cells outweighs the evidence from the disagreeing cells. Consistent pairings outside the main diagonal is thus for Q_A a strong cue against agreement.

In sum, the new test Q_A was found to be just a little less powerful than the other tests when k is small (k < 8). However, these tests proved to be far less specific. One possibility is that on a small percentage of the simulations, random pairings might have occurred, triggering these tests to significance but not Q_A . Hence, it is not clear whether the new test is truly less powerful or is simply more selective.

General Discussion

We examined the performance of five tests for determining whether two observers, each one classifying N observations into a set of k categories, agree or not above chance level. Monte Carlo simulations were used to compare the five tests with respect to their sensitivity to true agreement—a quality that translates into statistical power—and their specificity, that is, their insensitivity to anything but true agreement. As could be expected (see for instance Cohen, 1960), the standard $χ^{2}$ statistic on contingency tables, although not insensitive to true agreement, ranks well behind the other tests for power as well as for specificity. The κ-to-z tests (Cohen’s $z_{κ 1}$ , Fleiss et al.’s $z_{κ 2}$ ) and $z_{Σ 2}$ perform best on power, but they are impaired by their poor specificity, a problem that does not affect the ratio-based Q_A test.

The statistics described here provide a nice set of tools to quantify agreement and assess its significance. Confidence intervals are also defined, allowing for easy comparisons between studies. The only limitation of the present statistics is that they are all biased downward (and the bias is important for k≥ 10 and small N). Future work should try to find the correction to this limitation. In the meanwhile, the present statistics should not be used to detect significant disagreement.

Our results show that a significant test on kappa can indicate that agreement is strongly present in a few categories or weakly present in all categories (or a continuum between these extremes). On the other hand, a significant Q_A test only indicates that agreement is present in all the categories. The simulations showed that the Q_A test is the only test that is sensitive to agreement within all categories, not just a few, as seen by comparing the markedly different results in the Random excluding coincident pairings with the Random with possible coincident pairings simulations.

What does distinguish the κ-to-z and sum-of-z statistics, on the one hand, and Q_A , on the other, and wherefrom does the latter earn its high specificity? The uniqueness of Q_A comes from the fact that the whole matrix of agreement is used. Kappa-based and sum-of-z tests only exploit the main diagonal. Hence, they implicitly assume that information outside it is uninformative, which is obviously false, as we discussed with the first two examples.

Footnotes

Appendix A

Appendix B

Appendix C

Acknowledgements

We would like to thank Bradley Harding for his comments on an earlier version of this article.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

References

Brennan

R. L.

Prediger

D. J.

(1981). Coefficient kappa: Some uses, misuses, and alternatives. Educational and Psychological Measurement, 41, 687-699. doi:10.1177/001316448104100307

Cohen

(1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20, 37-46. doi:10.1177/001316446002000104

Everitt

B. S.

(1968). Moments of the statistics Kappa and weighted Kappa. British Journal of Mathematical and Statistical Psychology, 21, 97-103. doi:10.1111/j.2044-8317.1968.tb00400.x

Fleiss

J. L.

Cohen

Everitt

B. S.

(1969). Large sample standard errors of kappa and weighted kappa. Psychological Bulletin, 72, 323-327. doi:10.1037/h0028106

Forbes

Evans

Hastings

Peacock

(2010). Statistical distributions. New York, NY: Wiley.

Harding

Tremblay

Cousineau

(2014). Standard errors: A review and evaluation of standard error estimators using Monte Carlo simulations. Quantitative Methods for Psychology, 10, 107-123.

Kaas

Buhrman

J. M.

(1980). Mean, median and mode in binomial distributions. Statistica Neerlandica, 34, 13-18. doi:10.1111/j.1467-9574.1980.tb00681.x

Kraemer

H. C.

Periyakoil

V. S.

Noda

(2002). Kappa coefficients in medical research. Statistics in Medicine, 21, 2109-2129. doi:10.1002/sim.1180

Landis

J. R.

Koch

G. G.

(1977). The measurement of observer agreement for categorical data. Biometrics, 33, 159-174. doi:10.2307/2529310

10.

Stouffer

S. A.

Suchman

E. A.

De Vinney

L. C.

Star