FROM NULL HYPOTHESIS TESTING TO ESTIMATING EFFECT SIZE

Prevailing research traditions in management scholarship have led to a focus on exploring the nature of the association between independent and dependent variables. As part of this focus, and albeit it is less common in current research, scholars often conflate p-values as a measure of both the existence and the strength of a relationship. However, p-values do not yield information on the strength of a relationship.

Scholars face limitations when using p-values as the sole indicator of associations between variables. First, p-values are dependent on the sample size. One general property is that p-values tend to decrease with increases in sample size. Within very large samples, finding coefficient estimates with low p-values is often of little surprise (Sullivan & Feinn, Reference Sullivan and Feinn2012). Second, a p-value does not measure the strength of a relationship. If X ₁ and X ₂ are independent variables and they have p-values of 0.06 and 0.03, respectively, in their relationship with the dependent variable Y, it does not mean that X ₁ has a larger effect size or a more substantive impact on Y than X ₂. Third, an independent variable with a low p-value might not have a substantive influence on the dependent variable. In other words, we can have confidence that the observed association exists due to the demonstrated low p-value, but any movement on the independent variable might only lead to trivial changes on the dependent variable. Effect size analysis helps identify the level of change or what we can call the practical significance.

Perhaps most troubling from the focus on p is the emergence of the notorious practice of p-hacking. With a greater availability of large datasets and increasing ease of running increasing sophisticated models, p-hacking has become more common (Baum & Bromiley, Reference Baum and Bromiley2019). Understandably journal editors and researchers are concerned about the problem as p-hacking leads to practices such as opportunistic modifications to models, or limited reporting of models. As an example, a researcher can play with various combinations of control variables to identify the specification that moves the independent variable of interest from being above p = 0.05 to being below p = 0.05.

Increasingly p-hacking has been identified as an important and common problem in management research (Baum & Bromiley, Reference Baum and Bromiley2019). It can be damaging to the stability and replicability of findings (Lewin et al., Reference Lewin, Chiu, Fey, Levine, Mcdermott, Murmann and Tsang2016). Unfortunately, it is difficult to know whether p-hacking has occurred in a specific paper. One advantage of effect size measurements is that they are not prone to change with increases or decreases with sample size. By promoting both reporting of statistical significance and practical significance (effect size) of research, we can reduce the stress that authors face to find that elusive p < 0.05, as even that p-value might not have a large effect size. As a result, reporting effect size might help decrease the prevalence of p-hacking.

Recent management literature has identified well these aforementioned limitations in the use and interpretation of p-values (Meyer, Van Witteloostuijn, & Beugelsdijk, Reference Meyer, Van Witteloostuijn and Beugelsdijk2017). Alongside this development has been a push to gain stronger causal inference in research design. An increasing number of management scholars are now implementing research designs that use experimental methods such as quasi-experiments or randomized controlled trials. In addition, there has been increasing attention given to matched pairs or matched samples analyses, in which the variable(s) of interest can be isolated more effectively and then explored using difference-in-difference analytical techniques, for example. These techniques align strongly with the dictates of establishing evidence in support of causality, more so than can be obtained by using lagged variables in multivariate analysis. Yet, even though we have seen these changes in research designs and methods, we still see a substantial body of researchers using cross-sectional data either obtained via primary sources such as in-person or paper (online) mediated surveys, or from secondary sources such as archival databases.

The advancements in research methods to more acutely tease-out causal relationships have permitted researchers to design and report research more effectively. Clearly, p-values still carry important information for good statistical inference; however, we contend that the next stage of improvement in reporting on one's econometric analyses should be to calculate and communicate effect sizes. Effect size reporting should ideally become an essential part of the communication about a study's empirical results.

Importantly, we note that this consideration of effect sizes moves beyond the treatment effect considerations we discussed briefly above, where researchers attempt to show causality. By this, we mean that even within the consideration that management research has been making advances in terms of the methods by which to develop and show consistent and reliable empirical evidence that A causes B, it has been doing so within the confines of ruling out, so to speak, alternative explanations, such that C or D or E cannot be advanced as reasonable alternative causes of B. As much has been written about better causal inference, we focus our attention on the next steps of improvements in the research process.

Put another way, the starting point in our discussion is that we assume that the researcher has utilized up-to-date, accepted techniques for establishing reasonable causal inference for A to B. Given that the causal inference is acceptable, we then move to the importance of how to determine, report, describe, and depict the effect size of a relationship as an aid to understanding if the relationship has a large enough substantive impact to make a useful contribution to management theory and practice. Effect size is a useful complement to causal inference as it helps to show that an issue of interest has a large enough influence to be of interest to scholars and practitioners.

In terms of the background for some of our claims, the fields of psychology and education were early movers in the Social Sciences for promoting reporting effect sizes when presenting empirical results. Educational and Psychological Measurement made its first call for reporting effect size in 1994 (Thompson, Bhatt, Lazarus, Cullen, Baranowski, & Baranowski, Reference Thompson, Bhatt, Lazarus, Cullen, Baranowski and Baranowski1994). Furthermore, the American Psychological Association's task force on statistical inference followed up its initial 1994 recommendations in 1999 by stating that researchers should ‘always present effect sizes for primary outcomes … interval estimates should be given for any effect sizes involving principal outcomes’ (Wilkinson, Reference Wilkinson1999: 599). The APA has continued to stress the importance of reporting effect size (APA, 2020). The American Educational Research Association (2006) has also advocated the reporting of effect size. It is, however, worth noting that while calls for reporting effect size in psychology and educational research have increased substantially, effect sizes have not been as consistently reported and well interpreted even in psychology and education articles as desirable. Meanwhile, journals such as the Journal of Applied Psychology, Educational and Psychological Measurement, and Psychological Science have been early movers in requiring reporting effect size (Kelley & Preacher, Reference Kelley and Preacher2012).

Management journals have not been opaque to issues around effect size reporting when interpreting the managerial meaning of results. For example, Bettis, Ethiraj, Gambardella, Helfat, and Mitchell (Reference Bettis, Ethiraj, Gambardella, Helfat and Mitchell2016) in the Strategic Management Journal, Ellis (Reference Ellis2010) in the Journal of International Business Studies, the guide for authors for Organizational Behavior and Human Decision Processes (OBHDP Editors, 2021), and Lewin et al. (Reference Lewin, Chiu, Fey, Levine, Mcdermott, Murmann and Tsang2016) in Management and Organization Review have advocated reporting effect size. However, a clear guide with sufficient direction on when and especially how to report effect size has been lacking in management research. This has led to the situation where management research has fallen behind other Social Science disciplines in terms of informing audiences about the substantive meaning of results within its empirical research. Critically, without a sufficient analysis or consistent reporting of effect size, management research has not been able to identify fully the practical implications of its research on individuals and on organizations. As such, in the management field there is a lack of a clear understanding of where and how advice can be derived for practitioners. Studies fail to point to where practitioners could focus their efforts to improve the performance of objectives of interest as predicated upon an understanding of the nature of causality and the strength of a relationship. Focusing on effectively identifying and reporting effect sizes can help to rectify this persistent shortcoming in management research.

A recent anecdote helps illustrate how potential is currently being realized in internal research departments in large organizations. In the late 2010s, at the American Economic Association Meetings, Amazon was one of the most prolific recruiters of new PhD economists. The reason for Amazon's recruiting drive was that PhD economists have developed empirical skills by which they were not only able to develop models with strong causal inference, but they were also able to develop predictions about how strongly an independent variable X could affect a dependent variable Y. With good causal inference, and with good methods for identifying and reporting effect sizes, measurements of the effectiveness of marketing campaigns or other such initiatives could be distilled into a set of defensible numbers based on the application of contemporary econometric techniques to the very large data that Amazon possessed. We need to ensure that management scholars are equipped with and can use analytical tools to explore substantive impacts of manipulations of X to Y, such as what Amazon and other firms are doing with their economist PhD recruits.

MEASURING EFFECT SIZE

A core element to the measurement of effect size is the estimated coefficient in a regression estimation. In a linear regression, coefficient estimates represent the predicted change in the dependent variable corresponding to a unit change in the independent variable, ceteris paribus. For example, in a study of Standard & Poor's 1500 firms in the 1992–2006 period, Dezsö and Ross (Reference Dezsö and Ross2012) find that ‘… [a female member sits] in top management leads to an increase of $42 million in firm value…’. Effectively, this reporting of the substantive nature of the effect size of female representation provides a clear interpretation of the outcome in terms of firm value.

When it comes to non-linear relationships, the calculation is more complicated. Sampson (Reference Sampson2007: 377) employed a negative binomial model to identify a U-shaped relationship between a firm's technological diversity as measured by its alliances, and its innovative outcomes, as measured by its patents. In the reporting of the effect size of results, we find the statement that ‘… moderately diverse alliances contribute over 13 times more than alliances with very low diversity … and over 3 times more than alliances with very high diversity…’.

Sometimes, researchers can use mean differences between groups to compare effect sizes. David Card, a Nobel Laureate in Economics in 2021, reported in his classic study on minimum wage and employment that: ‘… the rise in the California minimum wage reduced the fraction of California worker earning $3.35–4.24 per hour by some 5 percentage points … [and] raised the wage 15% for 5% of California workers’ (Card, Reference Card1992: 42).

These three examples clearly illustrate how researchers can provide an indication of the practical significance of their research. This communication can aid managers and policymakers in the design of their organizations or programs for developing evidence-based initiatives. The reporting found in the above comes from calculations based on the substitution of coefficient estimates into the relevant regression equation to determine the prediction of the dependent variable, given the relevant unit changes in the independent variables. We note that this is an important way of reporting effect sizes in research which we recommend. It is what we identify as describing the effect size in words, but we also want to draw attention to other ways in which research has developed measures for effect size.

Table 1 depicts our summary of measures of effect size. We note there are over 40 measures of effect size, which challenges researchers to know which is the correct one to use in a particular circumstance. Table 1 aids this choice by identifying the effect size measure we recommend according to the type of commonly used statistical analysis employed in the research. Table 1 provides our effect size-related recommendations for numerous commonly used statistical analyses in management research. These effect size measures are developed based on the statistical attributes of the sample used in an analysis and its corresponding statistics. To help convey the meaning of different effect size measures, we identify what constitutes a small, medium, and large effect size for each measure. Table 1 is rich in information, making it a key reference for management researchers. We describe several specific areas in Table 1.

Table 1. Examples of commonly used effect size measures

As shown in Table 1, researchers have used estimated mean differences (Nakagawa & Cuthill, Reference Nakagawa and Cuthill2007; Rosenthal, Cooper, & Hedges, Reference Rosenthal, Cooper and Hedges1994), correlations (Cohen, Reference Cohen1992), variances (Fern & Monroe, Reference Fern and Monroe1996), and regression estimates (Cohen, Reference Cohen1988; Sullivan & Feinn, Reference Sullivan and Feinn2012) to measure and report effect size. Accordingly, a clear implication of the information in Table 1 is that researchers should choose a measure of effect size according to their research design and the statistics that are available.

As an example, researchers can use Cohen's d as an effect size measure when the independent variable is categorical and the dependent variable is continuous. Pearson's r can be used as an effect size measure when both the independent variable and the dependent variable are continuous.

When considering effect sizes, a natural question to ask is what constitutes a large, medium, and small effect size. Cohen was one of the earliest scholars to develop the concept of effect size. His work remains important to the present day. Among Cohen's (Reference Cohen1988, Reference Cohen1992) impression-based guidelines for different types of statistics, his recommendations for classifications of effect sizes for mean differences and correlations are especially widely cited today. For a comparison of mean differences, effect sizes between 0.20 and 0.49 are said to be small, effect sizes between 0.50 and 0.79 are medium, and effect sizes of 0.80 or greater are large. For a Pearson r, a correlation of 0.10–0.29 is a small effect size, a Pearson r value of 0.30–0.49 is classified as a medium effect size, and a Pearson r value of 0.50 or greater is a large effect size.

To provide some context of what these different effect sizes mean, Cohen (Reference Cohen1988) suggested that an effect size of 0.2 would be yielded by comparing the height of 15- and 16-year-old girls in the US. In other words, the effect size is not very large or easy to observe. He further suggested that unlike the 15- and 16-year-old comparison, a phenomenon leading to a 0.50 (medium) effect size should be visible to the naked eye. Cohen also warned he was trying to provide approximate guidelines that could be useful across different behavioral sciences. However, interpretations of effect sizes should also consider the limitations in using specific cut-offs to identify the magnitude of an effect size.

Cohen (Reference Cohen1988) recommended that researchers consider the levels that have been observed and deemed reasonable in past studies in a specific area, as an additional point for identifying the size of an effect size. Relatedly, studies such as Sun, Pan, and Wang (Reference Sun, Pan and Wang2010) warn against a blind reliance on effect size cut-offs. Just like the critique related to p-values that the difference between p = 0.05 and p = 0.06 is just 0.01 and should not substantially change the interpretation of a statistical analysis, effect sizes should be considered as a continual evolution from small to large with cut-offs providing guidance.

This point connects to the reasons why Cohen (Reference Cohen1988: 352) offered benchmark guidelines. He did so ‘because they were needed in a research climate characterized by a neglect of attention to issues of [effect size] magnitude’. Such a climate persists today given that effect sizes are not discussed in most published research. Thus, the recommended cut-offs for effect size continue to serve a useful function, but the context in which the research was conducted should be considered in their application.

It is useful to consider how well Cohen's (Reference Cohen1988, Reference Cohen1992) cut-offs apply to recent research. Based on a review of two meta-analyses in psychology, Hemphill (Reference Hemphill2003) found that one-third of the correlations in the meta-analysis were less than 0.20, one-third were between 0.20 and 0.30, and one-third were greater than 0.30. Hemphill (Reference Hemphill2003) proposed accordingly these three ranges be recast as small, medium, and large effect sizes for correlations, which are similar but slightly lower thresholds than those recommended by Cohen (Reference Cohen1988, Reference Cohen1992). More recently, Gignac and Szodorai (Reference Gignac and Szodorai2016) found that only 2.7% of correlations were 0.50 or larger, which suggests that this previous cut-off of 0.50 (Cohen, Reference Cohen1988, Reference Cohen1992) seems to be too high for a large effect size. As such, they suggest that a small effect size should be 0.10–0.19, a medium effect size should be 0.20–0.29, and a large effect size should be 0.30 or larger. Given the realities of observed inter-item correlations found in most management research, the recommendations for the qualitative categorization of effect size by Gignac and Szodorai (Reference Gignac and Szodorai2016) have the most substantive and analytical traction.

Next, Acock (Reference Acock2014: 272) suggested that standardized regression coefficients can be considered at similarly corresponding levels as those we have discussed for Pearson's correlations. More specifically, he suggests that a standardized β less than 0.2 should be considered as a weak association between the independent and dependent variables, whereas a β that is greater than 0.2 but less than 0.5 be considered as moderate, and a β greater than 0.5 should be considered a strong effect.

It is worth noting that caution should be exerted when considering β as a measure of effect size, as various factors affect the standard deviation of a variable and standard deviations are used to calculate a β. A common measure of the collective effect size of all the explanatory variables (the global effect size) in a multiple regression is Cohen's f ² which is R ²/(1 − R ²). Cohen (Reference Cohen1992) states that small effect sizes for Cohen's f ² are between 0.02 and 0.14; medium are between 0.15 and 0.34; and large effect sizes are greater than or equal to 0.35. In their 30-year-observation review of categorical moderators, Aguinis, Beaty, Boik, and Pierce (Reference Aguinis, Beaty, Boik and Pierce2005) found that Cohen's f ² averaged only 0.002. However, they found that 72% of the moderator tests had a power of 0.80 or more, which provides some evidence that smaller effect sizes for categorical moderators may be reasonable to consider.

Cohen's f ² can also be modified to measure the effect size of an independent variable of interest (see Table 1 for details). This is called Cohen's incremental f ² and has the formula of: f ² = $( R_{{\rm included}}^2 -R_{{\rm excluded}}^2$)/$( {1-R_{{\rm included}}^2 } )$. It has the same effect size cut-offs as mentioned for the global/regular Cohen's f ² above. This effects size measure can be helpful when building models incrementally or when one wants to identify the amount of information or the practical significance of a focal variable.

We want to point out that some effect size measures are standardized like β and f ². We normally advocate for the use of effect size measures that are standardized as they facilitate comparison across variables and studies. However, a limitation of such measures is that they can be less intuitive to understand due to not being presented in conventional units. Thus, it can be useful for researchers to also present non-standardized effect size measures to better understand results from a particular study. Furthermore, researchers are encouraged to report both local effect size measures that focus on the effect of one variable, such as by using Cohen's incremental f ² for key independent variables of interest, and global effect size measures like Cohen's f ² for an entire model.

BEST PRACTICES IN REPORTING EFFECT SIZE

We make two recommendations on how researchers should report effect sizes: (1) they should include a detailed textual description in the paper and (2) they should have a clear presentation in graphs. In the text of the paper, authors should report the estimated coefficients of a statistical regression, along with confidence intervals and/or standard errors of the coefficients. Using these statistics, researchers can explicitly describe the effect size, compare in qualitative terms (small, medium, large), and explain the substantive impact in words (Bettis et al., Reference Bettis, Ethiraj, Gambardella, Helfat and Mitchell2016; Ellis, Reference Ellis2010). Textual descriptions not only communicate the academic findings of a study but also let readers associate academic findings with real-world experience.

A recent paper published in MOR by Li-Ying, Zhang, and Long (Reference Li-Ying, Zhang and Long2018) is a good example of the implementation of this practice. Their research concerns how hidden knowledge facilitators, about whom a firm's employees may be unaware, can promote knowledge sharing within a firm. They report the estimated coefficients of the focal independent variable, the quantity of facilitators, to depict their findings about how facilitators help knowledge sharing. Thus, after establishing evidence that this relationship is reflected in their data and analyses, they go further to test the moderating effect of the quality of facilitators and report on page 808 that: ‘…The [standard mean difference] effect size for the moderating effect is -0.156 (with an estimation of [-0.209, -0.10] at the 95% confidence interval) …’. With this effect size analysis, the authors announce a failure to finding a moderating effect.

Gupta and Misangyi (Reference Gupta and Misangyi2018: 41) provide a good example of explaining effect size in words when they write, ‘Low (−1. SD) vs. high (+1. SD) values of recent international diversification actions by charismatic CEOs predicted a difference of 3% (2% for narcissistic CEOs) change in the DV, which, for a company with median sales of all firms in our sample, represents a difference of $300 million ($200 million for narcissistic CEOs) in revenues from international business segments.’ We also encourage researchers to report conventional effect size measures (e.g., Cohen's d, η ², pseudo-R ², and ω ²) as depicted in detail in Table 1. These conventional measures should be discussed and interpreted in plain, easily accessible (non-technical) language. Song, Liu, Wang, Lanaj, Johnson, and Shi (Reference Song, Liu, Wang, Lanaj, Johnson and Shi2018) provide a good example of this practice in their study of how prosocial intervention on employees can reduce perceived customer mistreatments. They use a pseudo-R ² to measure the effect sizes from the estimated coefficients in the model and report on page 1005: ‘Predictors … accounted for 13% of the within-person variance in employee daily experience of customer mistreatment, 16% in afternoon negative mood, 24% in evening rumination ….’ Moreover, they use Cohen's d and ω ² for cross validation. By providing rich information and by using multiple effect size measures, the authors show the power of their field experiments when they compare their findings with conventional standards and previous studies.

Finally, we suggest that authors indicate if the effect size corresponds to a small, medium, or large scale in established classifications. Song et al. (Reference Song, Liu, Wang, Lanaj, Johnson and Shi2018) again provide a good example as they refer to conventional cut-offs when reporting effect size measures: ‘…ω ² = .10, … [which] indicates a medium effect size (typically ranging from 0.06 to 0.15) for the repeated-measure design’.

Next, the visualization of effect size is becoming an increasingly common practice in the reporting of results in research in management and organizations. Visualization in reporting effect size provides a graphical description, and it helps one to quickly understand what is being hypothesized and observed. Furthermore, it enables readers to see beyond a conceptual model and more clearly recognize the substantive impact of the key findings of a study.

An important initial consideration before constructing a visual representation of the effect sizes of an analysis is to supplement information on point estimates with reported standard errors to identify confidence intervals for non-standardized regression coefficients. Reporting confidence intervals explicitly allows us to understand the level of precision in coefficient estimates. A common standard is to use a 95% confidence interval, which captures values within approximately ±2 times the standard error. When reporting a confidence interval, it is important to identify if the confidence interval includes 0. If it does, then we cannot confidently state that we have a positive or negative relationship, or even a relationship at all. Alternatively, if we are plotting the effects of two positive or negative coefficients and we believe these to be different in value from one another, we do not want to see overlap in the confidence intervals as projected for both. We will see this point more clearly in the example we present later in this section.

Next, a good presentation using graphs also necessitates a consideration of the units that are used, and the quantities or scales depicted. Furthermore, it provides a platform for researchers to demonstrate their creativity in reporting their findings. That said, we still believe that extant papers infrequently use graphical descriptions on effect size.

Chatterji, Delecourt, Hasan, and Koning (Reference Chatterji, Delecourt, Hasan and Koning2019) provide a good example in their research on whether start-up firms can benefit from receiving consulting advice from established entrepreneurs. They find start-ups can have an additional size growth rate of 28% as measured by the number of employees, and a 10% lower chance of failure, if firms receive advice from entrepreneurs (as compared to those who do not receive advice). The authors visually present the effect size of the main effects, which is the benefit of receiving entrepreneurial advice as related to the growth and survival of start-ups. They divide their presentation of effect sizes according to whether the firm is an MBA start-up or a non-MBA start-up (see figure 2 of their paper).

It is noteworthy that in this figure they provide point estimates with confidence intervals, which is an emerging but not common standard in graphical representations. As reprinted herein, Chatterji et al. (Reference Chatterji, Delecourt, Hasan and Koning2019) is a best practice example of how to visualize effect sizes, which we recommend MOR authors to follow. Another good example is Vanneste and Gulati (Reference Vanneste and Gulati2021) in their study of a firm's use of trust in coping with economic downturns. The authors provide graphs of effect sizes (figure 3 of their paper) across regression models for the illustration of the results of their four hypotheses. They include point estimates and confidence interval depictions, which provide an immediate and clear visualization of the effect sizes for readers.

Within MOR, we note several relevant examples of effect size depictions. Chen, Dai, and Li (Reference Chen, Dai and Li2019) provide a detailed discussion of how a firm's innovation output is associated with its affiliated R&D consortia. The authors used standard errors and confidence intervals in their discussion of the main effects and moderating effects. Prashantham, Zhou, and Dhanaraj (Reference Prashantham, Zhou and Dhanaraj2020) report a global effect size for a structural equation model and discuss effect sizes for quantitative empirical articles when reviewing the literature. Both steps are good practices that help readers to interpret their study and past research with accuracy. Another example is Hu, Zhang, and Yao (Reference Hu, Zhang and Yao2018) who study the labor productivity of family firms. Their figures 1–3 provide good examples of how to plot the effects of interaction terms, as related to effect size.

A CHECKLIST FOR REPORTING EFFECT SIZE

We strongly recommend authors to use both textual descriptions and visual aids to report and interpret effect sizes. We provide two checklists as guidance on how to implement this practice.

1. I. When reporting the effect size measures and discussing effect sizes in a textual description, researchers should:

   1. (1) provide a description of the estimated coefficients;

   2. (2) provide confidence intervals or report the standard deviations;

   3. (3) provide measures of effect sizes, which can include a comparison of the effect sizes observed in relation to those found in previous related studies;

   4. (4) when appropriate, make reference to established norms for cut-offs to identify if an effect size is large, medium, or small;

   5. (5) provide an interpretation of the effect size in words such as an interpretation of the economic meaning, if any; and

   6. (6) report both the effect of key independent variables of interest using effect size measures such as Cohen's incremental f ², and the global effect size of the entire model using measures such as Cohen's regular f ².

2. II. When presenting a visualization of effect size, researchers should:

   1. (1) provide point estimates of effect sizes;

   2. (2) provide a visual depiction of confidence intervals around the point estimates;

   3. (3) develop fitted lines that link points and show the trend of estimates for the dependent variable along values of the independent variable;

   4. (4) all values within the figure should be within the ranges of the independent and dependent variables as found in the data;

   5. (5) the plots should be developed based on fitted values of the dependent variable. The calculation uses the estimated regression equation, where specific values in the range of the key independent variable are used, alongside the mean values of all other variables in the regression equation (Makino & Delios, Reference Makino and Delios1996: 918); and

   6. (6) identify the confidence intervals around critical isolated point estimates found in the plots, resulting in a figure mirroring our reprint of Chatterji et al. (Reference Chatterji, Delecourt, Hasan and Koning2019), as shown in Figure 1.

Figure 1. Examples of plots of effect size as reproduced from Chatterji et al. (Reference Chatterji, Delecourt, Hasan and Koning2019, figures 1 and 2)

These two checklists refer to ideal presentations of effect size, which should become part of standard practice in reporting. Admittedly, space can be scarce in journals. Researchers must balance the need to have effect size discussions in the main text of the paper versus the inclusion of other material, where at times effect size visuals might be presented in appendices. Kaiser, Kongsted, Laursen, and Ejsing (Reference Kaiser, Kongsted, Laursen and Ejsing2018) and Asgari, Tandon, Singh, and Mitchell (Reference Asgari, Tandon, Singh and Mitchell2018) recommend publishing graphs and critical statistics in appendices that are available online. In so doing, space constraints become less of a valid excuse for not reporting effect sizes. Researchers are thus empowered to provide a complete set of textual and graphical descriptions of effect sizes.