11/03/2024
This annual reporting season has seen the highest number of strikes (more than 25% votes against) for remuneration report votes since the introduction of the twostrikes rule in 2011. Company directors and proxy advisors alike have been debating on the causes of this surge in protest votes. Some have argued this change in sentiment to be temporary and partially the result of cyclical economic conditions. Others have commented that it may reflect a longerterm shift in voting patterns, while others yet have argued it is due to a combination of these reasons.
In light of these recordbreaking shareholder voting trends, we have undertake research to consider some of the underlying factors. Consistent with prior years, we have conducted an indepth analysis to determine what factors may be behind voters’ actions, appending our latest data and analyses onto our longrunning archive of annual studies. We have conducted a more detailed analysis this year, to address the voting anomalies observed.
Our previous articles (see HERE, HERE, HERE, HERE, HERE and HERE) involved similar analyses. This year’s analysis builds on prior studies with the addition of the ASX 300 companies’ 2023 annual general meetings (AGMs) and performance data.
Analysis
We explored potential associations between the proportion of oppositional votes against remuneration report proposals, total shareholder return (TSR), total assets, return on equity (ROE), business sectors (as per the Global Industry Classification Standard), and relative TSR figures for each sector. We looked at all ASX 300 companies over a time horizon spanning 12 years, from 2012 to 2023 inclusive.
By taking account of these factors, we examined whether the voting trends observed may have been due to company performancerelated issues, and whether the patterns were dependent upon company size or sector.
We follow the convention of a 5% significance level, meaning that we tolerate a risk of committing Type I errors (false positives) of up to 5%. When reporting Pvalues for hypothesis tests, we use the following significance codes:
0.05 ≤ P < 0.1 
0.01 ≤ P < 0.05 
0.001 ≤ P < 0.01 
P < 0.001 
. 
* 
** 
*** 
The Methodology section at the end of this article provides detail on the statistical models and procedures used.
From our analysis, we have drawn the following conclusions:
 Votes against sayonpay resolutions were substantially higher this year relative to prior years (2012 to 2022), and this increase was statistically significant.
 Disparities in mean percentage votes were observed between different GICS sectors, but none of these were statistically significant.
 TSR had a statistically significant negative effect on oppositional sayonpay voting, which aligns with what we observed last year. This indicates that betterperforming companies tended to face less opposition to their remuneration report proposals.
 The negative impact of TSR upon oppositional voting was significantly modulated by ROE and the relative TSR rankings by sector.
 Total assets exerted a statistically significant positive effect on oppositional voting, indicating that larger companies faced more opposition.
 Neither ROE nor relative TSR by sector were found to have any statistically significant influence on voting outcomes, outside of their interactions with TSR.
Over twelve years of data, it has been consistently observed that ROE – despite being a logical performance measure – seldom impacted shareholder voting outcomes on remuneration report proposals, whenever TSR had already been considered. Our conclusions differ slightly this year, as we have now accounted for its interaction with other economic factors.
Longitudinal trends
This annual reporting period saw an anomalous surge in oppositional votes, as illustrated in Figure 1. On average, votes against remuneration reports have increased from 7.69% last year to 12.56% this year. Over the past decade, this figure tended to hover between 6% and 8%. This surge in shareholder dissent is also reflected in a nearly twofold corresponding increase in strikes from 22 strikes in 2022, to a record high of 41 strikes in 2023. This was the highest tally of strikes since the introduction of the twostrikes rule in 2011.
Figure 1: Mean % votes against remuneration reports for each year (2012 to 2023) for ASX 300 companies.
By performing multiple comparisons at a 5% familywise error rate (FWER) (refer to Methodology for details), we concluded that – on average – oppositional votes for 2023 differed significantly from all prior years. Otherwise, the years did not differ significantly. Table 1 displays the estimated differences in means for each pair of years with P < 0.1, along with the lower and upper bounds of the 95% confidence intervals, the Pvalues, and their significance codes.
Table 1: Pairwise comparisons of means of percentage votes for each pair of years (2012 to 2023), with P < 0.1. The only pairs of years with P < 0.1 were the ones involving 2023.
Comparison 
Difference 
Lower 95% Conf. Bound 
Upper 95% Conf. Bound 
PValue 
Significance 
20232012 
4.51% 
0.72% 
8.30% 
0.00580803 
** 
20232013 
6.38% 
2.62% 
10.14% 
2.09627E06 
*** 
20232014 
5.49% 
1.76% 
9.23% 
0.000102288 
*** 
20232015 
5.59% 
1.93% 
9.26% 
4.15044E05 
*** 
20232016 
6.42% 
2.83% 
10.01% 
3.60016E07 
*** 
20232017 
6.67% 
3.14% 
10.20% 
4.64774E08 
*** 
20232018 
4.90% 
1.40% 
8.40% 
0.000298676 
*** 
20232019 
5.26% 
1.71% 
8.81% 
8.60195E05 
*** 
20232020 
4.08% 
0.55% 
7.62% 
0.008817472 
** 
20232021 
5.23% 
1.80% 
8.66% 
4.08586E05 
*** 
20232022 
4.88% 
1.32% 
8.43% 
0.000459885 
*** 
Trends by sector
On average, the business sector with the highest level of sayonpay protest votes was Utilities at 10.88%, while the sector with the least dissent was Real Estate at 5.94%. The mean oppositional votes for each sector are visualised in Figure 2. The means across sectors tended to hover around 8%. From another multiplecomparisons analysis (with a 5% FWER), we concluded that the observed differences in percentage votes between GICS sectors were not statistically significant.
Figure 2: Mean % votes against remuneration reports for each GICS sector, for all years (2012 to 2023).
Impact of financial factors
In a logistic regression analysis (detailed in Methodology), we considered the impact that the years, GICS sectors, TSR, and log total assets all had in influencing the level of protest votes in sayonpay proposals. We considered the influence of TSR through both raw percentages, and relative TSR (rTSR) percentile ranks compared to the rest of the company’s GICS sector in the same year. Logtransformed total asset values were used to improve the fit of the linear model (refer to Methodology for interpretation).
Within the model, TSR had a negative impact on voting outcomes, whereas the impacts of sectorwise rTSR and total assets were both positive. Interaction effects were also modelled; the two significant interaction terms were the TSRrTSR term and the TSRROE term. Both interaction effects were positively related to voting outcomes, within the context of the regression model. However, ROE as not found to be important for the model as a standalone independent variable, reflecting our findings in prior years. The results of the regression analysis, such as coefficient estimates and Pvalues, are given in Methodology.
In Figure 3, we consider the influence of TSR upon voting outcomes through each company’s annual rTSR percentile rank, comparing them to all other ASX 300 companies from the same year. The percentile ranks are grouped into deciles. The annual rTSR gives an idea of how each company’s TSR compared to the rest of the ASX 300 in that year. Figure 3 illustrates that protest votes tended to decline as company performance improved. ASX 300 firms in the highest rTSR decile experienced the least shareholder dissent in sayonpay, with a mean of 5.56% votes against. Conversely, companies in the lowest decile faced the greatest dissent, with a mean of 10.75% votes against.
Figure 3: Mean % votes against remuneration reports for each annual rTSR decile group (relative to ASX 300).
We now instead consider the impact of total assets upon voting outcomes, by again dividing the companies based upon which decile their total assets fall into, as shown in Figure 4. A clear positive trend is evident, with higher asset deciles being associated with more protest votes. The highest average number of sayonpay protests occurred in the top decile, at 9.69%. Conversely, the lowest mean number of protests occurred in the fifth decile, at 6.74%, while the bottom decile saw a similar mean of 6.91% protest votes.
Figure 4: Mean % votes against remuneration reports for each totalassets decile group (ASX 300, all years).
Conclusion
Our findings for this reporting period were mostly consistent with previous analyses.
TSR is negatively associated with protest votes to sayonpay. Weak TSR results cause increased shareholder dissent relative to strong TSR results.
Company size, as measured by total assets, is positively associated with protest votes. Larger companies – especially those in the top decile by total assets – face more difficulty in mitigating shareholder dissent against their remuneration report proposals, relative to smaller companies.
Good ROE results are overshadowed by TSR in determining voting outcomes, but the interaction between ROE and TSR was found to be an important factor.
This year, we can draw further conclusions from our analysis.
Regarding longitudinal trends, significant differences in voting outcomes were found between 2023 and all prior years (2012 to 2022), but not between the previous years. This reflects the anomalous surge in protest votes seen this year.
Some disparities in votes were found between the GICS sectors (combining all years), but these differences were not statistically significant.
Methodology
We analysed the impact of various economic factors upon sayonpay resolutions at AGMs held by ASX 300 firms, from 2012 to 2023 inclusive, using logistic regression. In addition, we compared voting outcomes for different years and GICS sectors using a multiple comparisons procedure (Tukey’s HSD test). For all hypothesis tests, a significance level of 5% was employed, using a familywise error rate (FWER) of 5% for multiple comparisons. The FWER is the chance of making at least one Type I error (false positive) in all pairwise comparisons. Company information and financial metrics were sourced from LSEG Workspace (formerly Refinitiv), while voting outcomes were sourced from Diligent Market Intelligence.
Our regression model was selected using a “best subsets” model selection process, which we employed to determine the optimal subset of independent variables and interaction effects (up to second order). The response variable was the percentage of votes against sayonpay proposals. We employed a logistic regression model, with the Logit link applied to the response variable:
Logit(PercentAgainst) = Log(PercentAgainst / (1 – PercentAgainst))
where Log is the natural logarithm, and PercentAgainst is the percentage of oppositional votes out of the total votes cast by shareholders. The Logit can be interpreted as the Log of the odds, or logodds, of oppositional voting. The Log and Logit functions both maintain the order of values, and hence the direction of inferred associations.
The initial candidate set of independent variables included: year; GICS sector; TSR; relative TSR using the companies’ GICS sectors in the same year as their peer groups; return on equity; and log total assets. The TSR was given by the oneyear figure ending on AGM date, whereas the ROE was given by the oneyear figure ending on the company’s financial year end. The log transformation was applied to total assets to linearise the relationship with the response variable.
After performing best subsets model selection, we arrived at the following optimal linear regression model using the adjusted Rsquared criterion for goodnessoffit:
VotesAgainst = β_{0} + β_{1}*2012 + β_{2}*2013 + β_{3}*2014 + β_{4}*2015 + β_{5}*2016 + β_{6}*2017 + β_{7}*2018 + β_{8}*2019 + β_{9}*2020 + β_{10}*2021 + β_{11}*2022 + β_{12}*Energy + β_{13}*Materials + β_{14}*Industrials + β_{15}*ConsumDisc + β_{16}*ConsumStap + β_{17}*Healthcare + β_{18}*Financials + β_{19}*InfoTech + β_{20}*Utilities + β_{21}*TSR + β_{22}*rTSR + β_{23}*LogAssets + β_{24}*rTSR*ROE + β_{25}*rTSR*LogAssets + β_{26}*TSR*rTSR + β_{27}*TSR*ROE + ε
where the years (20122022) and GICS sectors (Energy, Materials, etc.) were represented by dummy variables. The baseline year was 2023, while the baseline GICS sector was Real Estate. The initial candidate regression model, used as the starting point for the best subsets procedure, contained all the candidate independent variables, and all secondorder interaction terms not involving the year or GICS dummy variables.
The best subsets selection procedure involved identifying the optimal model for each number of predictors, ranging from one predictor up to a full model containing all 31 possible predictors (including interaction terms). For each model size, both the multiple Rsquared and the adjusted Rsquared were computed, as shown in Figure 5. The highest adjusted Rsquared occurred for the model with 27 predictor variables, which we hence selected as the optimal model.
Figure 5: Multiple Rsquared and adjusted Rsquared for each bestsubsets regression model size.
All regression models were fitted using a training data set, consisting of a random sample of 70% of all observations. This enabled an assessment of potential model overfit, by assessing the predictive performance of each model on the heldout 30% of data, as measured by the rootmeansquare prediction error (RMSPE). The RMSPE for each bestsubsets model size (from one predictor up to 31) is shown in Figure 6, indicating similar predictive accuracy for all models with more than two predictors.
Figure 6: Rootmeansquare prediction error for each bestsubsets regression model size.
The optimal regression model (fitted on the full data set) was found to be collectively significant (F = 8.797 on 27 and 2630 DF, and P < 0.001), with a multiple Rsquared of 8.283%. This Rsquared differs from the value shown in Figure 5, which corresponds to the Rsquared of the model fitted on the training data set, as opposed to the full data set. The coefficient estimates, standard errors, and Pvalues (with significance codes) are displayed in Table 2 below.
Table 2: Coefficient estimates and standard errors for independent variables and interaction terms in the optimal bestsubsets regression model, with Pvalues for the ttests of individual significance.
Variable 
Estimate 
Std. Error 
PValue 
Significance 
(Intercept) 
6.348464994 
0.61588712 
1.88358E24 
*** 
X2012 
0.274200622 
0.155400388 
0.077767536 
. 
X2013 
0.546993313 
0.154823959 
0.000417945 
*** 
X2014 
0.532873988 
0.153023209 
0.000505219 
*** 
X2015 
0.649527387 
0.151227069 
1.80991E05 
*** 
X2016 
0.65967126 
0.150796179 
1.26399E05 
*** 
X2017 
0.707533252 
0.146533146 
1.45456E06 
*** 
X2018 
0.509511523 
0.143113078 
0.000377137 
*** 
X2019 
0.438698614 
0.14696199 
0.002860832 
** 
X2020 
0.459477971 
0.145357245 
0.001590032 
** 
X2021 
0.457341796 
0.143803677 
0.001488238 
** 
X2022 
0.474894779 
0.145655478 
0.001126755 
** 
Energy 
0.315910678 
0.152281072 
0.03812776 
* 
Materials 
0.426122104 
0.106122854 
6.10148E05 
*** 
Industrials 
0.320025295 
0.123247011 
0.009467115 
** 
ConsumDisc 
0.45881641 
0.120364502 
0.000141072 
*** 
ConsumStap 
0.329058863 
0.160851201 
0.04088177 
* 
Healthcare 
0.672951479 
0.138904842 
1.34168E06 
*** 
Financials 
0.433482277 
0.111651498 
0.000105951 
*** 
InfoTech 
0.849616092 
0.154973699 
4.59686E08 
*** 
Utilities 
0.863555495 
0.289218982 
0.002854289 
** 
TSR 
1.072353296 
0.219297132 
1.06932E06 
*** 
rTSRsector 
0.549326354 
0.95992864 
0.567196456 

LogAssets 
0.144773201 
0.027480023 
1.48841E07 
*** 
rTSRsector:ROE 
0.301070398 
0.195709967 
0.124083394 

rTSRsector:LogAssets 
0.027953503 
0.045138781 
0.535786168 

TSR:rTSRsector 
1.052001594 
0.21877376 
1.60536E06 
*** 
TSR:ROE 
0.149033187 
0.053346083 
0.005248546 
** 
The raw data set consisted of 2680 observations. Of these, 8 had unavailable ROE data, one had unavailable total assets data (as well as ROE), while another 14 had zero values for oppositional votes, which caused issues with taking the Logit transformation. As a result, after necessary data cleaning, 2658 observations were examined in the regression analysis.
The 2023 data excluded companies with December financial year ends that had not yet held their 2023 AGM at the time of our analysis. This may cause some skews in the 2023 data that are currently unavoidable.
© Guerdon Associates 2024 Back to all articles