Crowdfunding: cross-country analysis

Table 8 The probit regression models results for Boomstarter and Kickstarter data for different countries for 0% or 100% collected.

The next step of our research was to run a probit regression model for 5 countries. In this model we used only variables, available on both Kickstarter and Boomstarter platforms. As it was mentioned when describing correlation matrix, there could be heteroscedasticity between the number of investors and the successfulness of the project. Moreover, when analyzing the descriptive statistics and differences between means, we also understood, that for all the projects, which collected 0% of their targeted fund, the number investors on Kickstarter equaled 0, which doesn't allow us to build regression with this variable. Thereafter, before starting running our regression model, we decided, that, as far as there are no significant difference for our research between unsuccessful projects and those, which collected at least something on Kickstarter and Boomstarter, we decided to run two probit regression models. The first one will estimate the factors influencing the probability of obtaining 100% financing rather than 0%, whilst the second one is to estimate factors influencing the probability of collecting 100% and more rather than less than 100%, but more than 0% and see if there are differences.

Firstly, we decided to run the regression model for the probability to collect 100% of targeted sum rather than nothing. The results of this regression may see in Table 8. We also want to remind that by running probit regression model we do not directly estimate the probability of obtaining 100% funding, but the change of our latent variable, which then we will estimate with average marginal effects to obtain the real influence degree of all factors on our percentage funded. What we obtain now is estimation of the latent variable changes and the side of effect (whether it is positive or negative) on our percentage funded. First important thing to take into consideration is that there is only linear relation between logarithm of the targeted fund of the project and our latent variable value for Russian crowdfunding platform. Moreover, in case of Russia, the coefficient of the targeted sum logarithm is negative. However, the opposite situation may be seen for other countries, placing their projects on Kickstarter platform. The situation is next: there is a positive relation between targeted sum logarithm and the value of our latent variable, however, if we take our non-linear variable, the situation is the opposite. When running average marginal analysis, we're going to understand the relation between all variables and the probability of obtaining 100% collected. For now, we know that targeted sum logarithm is significant, and we have a statistical evidence (the fact that we have relation between latent variable and targeted sum logarithm) that it is related to our dependent variable.

One more important moment is that all the categories, which we also included as control variables, are significant both in Russia and other countries. Thereafter, we'd better see the differences in coefficients. For Russian crowdfunded projects, all the categories have significant, but negative effect. As for the other countries, the effect is positive and also significant. However, in USA, the technological category has low, but negative coefficient. This may be the result of the fact that technological sphere is much more complicated, than Entertainment, for example, which has high enough and positive coefficient in a model. Thereafter, when making decision to invest, people are more likely to invest in things they may understand and use right away. Moreover, if we look at the Design category coefficient in Australia, we can see that it has the highest influence among all the categories in this country. Afterwards, we want to sum up, that the different categories have different influence for the project success. For now we see the statistical evidence that there is an influence of category on success, however, we still may not estimate the exact influence of all the variables on the probability for the project to succeed.

Next, we will look at the core measures of the goodness of fit of our regressions. As we may see in table 8 the pseudo R-squared estimate of our regressions is nearly similar for all countries, with main differences in Russia with slightly higher coefficient and in US with slightly lower one. Thereafter, after obtaining the results of these regressions, we'll see the values equal to 1 predicted by a model for a project which had 100% funded with 10%, 9%, 6%, 6%, 6% and 4% probability respectively for Russian, Australian, Canadian, British and American projects. The probability is rather low, however, the number of factors, included in models are not enough for understanding the project success terms. As to Hosmer-Lemeshov, or the goodness of fit test, the most important coefficient is the p-value of this test, as far as chi2 is just a squared difference between real and estimated values of dependent variable. Thereafter, returning to the p-value, we may say that we may reject the null hypothesis of the fact that the difference between the observed and predicted values is insignificant only for US model. As for other countries, the difference between observed and predicted values is significant on less than 1%, 62%, 66% and 81% confidence levels for Russian, Australian, Canadian and British models, respectively. Thereafter, we may conclude that this model fits US data perfectly, while it doesn't fit the Russian data. For other countries, the discrepancy of predicted values and observed data is rather high. The last estimate is the AUC, or the area under the ROC curve (Hernбndez-Orallo, 2013). The ROC-curves for the cross-country regressions may be seen on picture in Appendix 1. As we may see, these country-specific regression models may predict only up to 70% of positive observations (e. g. the dependent variable observations equal to 1). Thereafter, we may conclude, that these models are good enough as they may predict the successful outcome to a certain degree, however, they'd better to be more specified. For now, we are moving on to the description of average marginal effects for these regression models.

The average marginal effects for the first country-specific model may be seen in table 8. As we may see from the table, the logarithm of the targeted sum is significant on 95% confidence level and has a negative effect for the probability of collecting 100% of the sum, in case of Russia. This fact proves the result of the work “Guidelines for Successful Crowdfunding” written by Forbesa & Schaefera (2017) and the others see table 1 (summary of studies) in context of Russia. The same situation is for categories, they also are significant on 90% to 95% confidence level. As for the other countries, there are also nonlinear effects of targeted sum on the probability of our dependent variable to be equal to 1. Thereafter, for these two predictors we decided to make graphical interpretation in order to see the way they both relate to the probability of collecting the 100% of the targeted sum. As we may see on the picture in Appendix 2, the relations between the logarithm of the targeted sum and the probability of obtaining 100% fund are opposite for all countries to Russia. This is the result of the negative coefficient for Russian crowdfunding and the positive one for others. As we may also see that the projects in Russia has less than 50% probability to collect all the targeted sum if their targeted sum logarithm is more than 9. As for the other countries, for all of them if the targeted sum logarithm is equal or more than 7, thereafter, the probability of the project success, being placed on Kickstarter and run in one of these countries, is higher than 50%. However, in case of all other regressions, except for Russia, we need to look at the squared logarithm, as fat as it is also significant in these models. Thereafter, we may see the squared logarithms of the picture in Appendix 2.

The important thing to indicate is the fact that all the S-curves for the squared logarithms of targeted sums are similar to the Russian logarithm of targeted sum graph. Thereafter, we may conclude that for all countries except Russia, the prior variable to influence on the probability of the fund to be collected is the squared one. Thereafter, let us understand which targeted sum may be maximum for having at least 50% probability of success in these countries. As we may see on picture in Appendix 2, to raise the probability of success to at least 50% in Australia Canada and Great Britain, we need to set a squared logarithm of the targeted sum nearly 75 maximum, which equals to almost $6,000 of the targeted sum. As for the Design category, for example, this amount is quite enough for starting a small startup. Next, as for the United States, the maximum sum which may be achieved on Kickstarter by Americans with 50% probability is slightly higher and equals to $13,000. Thereafter, by setting this sum or the lower one, there is minimum 50% probability for US projects to be funded on Boomstarter. Now, we'd like to understand, which categories better to choose starting a project on Boomstarter and Kickstarter.

What is also interesting from these probit models, is that if the project on Russian crowdfunding platform, it will lessen its probability to collect the targeted amount for 15%, 14%, 20% and 8% if it will be from Design, Entertainment, Technology or Art category, respectively. Conversely, the probability to gain all 100% targeted sum will increase by 37% if the project will be of Design category and run in Australia. As for USA, the only category, which may lessen the chances of the project to collect all the targeted amount is Technological one. However, the difference will be quite low, the US founder will face only 1,5% risk of not collecting the 100% sum, if all other factors are equal, if placing Technological project on Kickstarter. As for Great Britain and Canada, the Design category is also a winning one there, as it leverages the chances to collect the targeted sum for 21 and 26% respectively. As for the Art category, the most successful project of this one are run in Australia and Great Britain. However, these results were counted for win or lose binary variable, thereafter, next we're going to build a regression model for the dependent variable equal to 1 if the project collects 100% of the targeted sum and equal to 0% if the project did not achieve the target, however, collected a certain amount of money.

Table 9 Marginal effects on the probability of 100% funding the project in 5 countries. Calculated as Average of Marginal Effects (AME)

Next, we estimated how different factors influence successfulness from country to country using probit regression for another dependent variable wherein 0 - goes for 0-99% funded and 1 for 100% or higher collected. The results of this regression may be seen in table 10. First of all, in comparison to previous cross-country regression, all countries do have negative link to the requested sum, the same happens if we take this variable squared.

Table 10 The probit regression models results for Boomstarter and Kickstarter data for different countries for 0-99% or 100% collected.

In the following case, number of investors is significant, this can be explained as in per_category2 we included projects with collection rate from 0% to 99,99%. There is positive relationship between the variables for all countries but in case of Russia a coefficient is much smaller. This can be explained by the difference in attitude to investing in chosen countries, as the mean number of a project's investors is lower due to the fact that the crowdfunding community in Russia is much less in comparison to developed countries.

Concerning the categories, the results varies between countries as well as in the previous probit, in Russia and the USA all categories are significant which confirms our hypothesis related to category dependency on success. As it can be seen from the table, Entertainment category is significant for all countries and has the highest coefficients in the conducted regression, moreover, they even increased after including projects with result between 0% and 100%. However, in Australia Design it is not the one of the highest influences, it is insignificant, the coefficient of Art is much lower as well. In case of technology the situation changed greatly, now it has negative effect on every country not only on USA.

Table 11 Marginal effects on the probability of 100% funding, not less than 100% for the project in 5 countries. Calculated as Average of Marginal Effects (AME)

As we may see from table 11 , the squared logarithm of the targeted goal for the project has negative coefficients in case of all countries, however, it is quite small, and equals to only 0.2%, 0.5%, 0.3%, 0.5% and 0.8% decrease in success probability in case of Australia, Canada, Great Britain, USA and Russia decrease in case of 1 unit increase in squared logarithm of the targeted sum for the project.

Moving on the squared logarithmic number of investors, the coefficients are negative for all countries with an exception for Russia, varying from 0.1% to 0.3% while in our county 0.1% with positive sign. That means, in Boomstarter collection rate increases on 0.1% if the squared logarithmic number of investors increases on 1 point.

Turning to the categories, authors who have chosen design category have lower probability of success in each country, the coefficients are 28.6%, 1.9%, 4.2%, 0.2%, 4.8% in Russia, Australia, Canada, Great Britain and USA respectively. The same happens with technologies, so that to moderate risks investors should avoid these categories.

The categories' coefficients are rather low in foreign countries, for the UK the highest influence have Entertainment and Art with 7.1% and 5.5%, while in USA 4.2% for Entertainment and -4.8% for design. In case of Russia, coefficients are negative for each category, moreover, they are the highest in comparison to the other countries from 23% to 28.6%, which indicates that the probability collect required sum on the russian platform is much lower if the other factors are equal and authors should probably consider foreign platforms to place a project.

Next, turning to the squared logarithm of the number of investors squared, the coefficients are negative for all countries, except Russia. The coefficients also too small, however, we still may them interpret as the ones, influencing change in the probability of obtaining the 100% of the targeted sum for a project. Thereafter, we may say, that, the probability of the project to succeed decreases for 0.1 to 0.3% when the author has the 1 unit greater logarithm of the number of people invested in the project page for all countries, except Russia, for which the probability increases for 1.4%.

Moving to the goodness of fit in our regressions. As we can see from the table above, the pseudo R-squared values are almost equal to all countries, approximately 70% plus with an exception for Russia with slightly more than 50%, and much bigger than in the previous cross-country regression in table 8. The results of the model predict well Australia, Canada, USA, Great Britain, but still in case of Russia some factors are omitted. Then, Hosmer-Lemeshov the results, the difference between the observed and predicted values is significant for all countries and fits the data perfectly.

However, there is a high possibility that we have omitted some factors influencing success rate in Russia, that is why we decided to study project's features authors traits in detail in the next part of our research. Now, we may turn to average marginal effects analysis for the Boomstarter regression model, the results of which may be seen in table 13. As we may see from the table, the coefficients for the targeted sum logarithms are very similar for all regressions, except the fourth one. This is due to the fact, that in fourth regression model, there is a squared logarithm of targeted sum, thereafter, the non-linear relation exists. As for other regressions, we may say, that increasing our targeted sum logarithm for 1, we're decreasing the probability to collect 100% targeted sum for 13% in case of the first regression with author-specific components in it, for 10% without author-specific and for 10% in case of the second regression with author-specific components and for dependent variable indicating whether the project collected at least something or nothing at all.

Table 12 The probit regression models results for Boomstarter data, with and without author-specific variables

As for the second regression without author-specific component, we may see that there is a squared relationship, thereafter, the coefficient, important for discussion is those for the squared variable. As we may see, the coefficient is rather low, and we'll get only 1% decrease in probability to fund our project for more than 0% if we raise our squared logarithm of targeted sum for 1 unit, which equals to nearly 2 rubles change. However, as far as our main objective is obtaining 100% funding, not at least something, we'd better go further. As for the last regression model, which is also important for us, due to the fact that by running it, we may understand the factors influencing the turning of the project from those, collecting at least something, to the one, which may collect 100% targeted sum. As we may see from the picture in Appendix 6, the squared logarithm of the targeted sum should be equal to nearly 80, in order to correspond to minimum 50% of the probability to obtain 100% funding, which corresponds to the nearly 8000 RUB. The number is not that big, however, as we understand, this and lower targeted sums may provide 50 and more percent of success probability. We consider that the result of such a low probability for high fund projects to succeed is the aftermath of the fact that the Boomstarter crowdfunding platform is still undeveloped and not as much popular as the Kickstarter one.

Next, we shall look at the author's biography length logarithm. As for the first regression, we may see that the coefficient is positive and small, thereafter, the influence is small. Thereafter, we may conclude, that no matter, if there are author-specific components in a model, the 1 unit increase in author's biography length logarithm provokes the 2% increase in the probability to gain 100% funding. As for the second regression model, there is a significant squared variable of biography length, thereafter, we'll pay attention mostly to it. As we may see on the pictures in Appendix 6, the longer the biography, the higher the probability of success (in this case, the probability to collect at least something) for two first regression models. From the table we may see, that the increase of the squared biography length logarithm by 1 unit, we increase the probability to obtain the funding of more than 0% for 0.1%, which is low, however, if we count it to symbols, it will be nearly 3 symbols, if we multiply them and our probability by 10, we obtain 30 symbols increasing the probability to 10%, which is the great plus, however, again, we just calculating it in the model, containing lots of specificators, which may lessen its value. As to the graphical results, which we may see in Appendix 6, there is an interesting tendency to describe. Despite, there is an upward trend for biography length logarithm and its square in two first regression models, the last regression results slightly differ from these ones. The last regression, both for author and project-specific variants, shows that the probability of success doesn't differ in case of biography length, as far as the graph is almost a straight line, which may be seen on pictures in Appendix 6. Thereafter, in case of these regressions, the biography length logarithm may be included as a constant.

Table 13 Marginal effects on the probability of 100% funding the project. Calculated as Average of Marginal Effects (AME)

Next, as to the age of the author, the average age, which is maximum for achieving 50% probability of success for a project, is 30 years. Thereafter, people, who is of age 30 are more likely to achieve the targeted sum, they put on Boomstarter. As for the number of followers, the mean number of the followers on the author's VK page ranges from 4 to 700,000 people, which has the positive upward trend to the success probability, which is ranging from 20% to 70%, it proves results of (Beier & Wagner, 2015) and the other authors see table 1 (summary of studies).

Moving on, we'll shortly describe the results of the binomial variables in our regression models. As a result, we may see that on average, if the project founder is female, thereafter, she has 6% lower probability to collect all the targeted sum if we talk about success/unsuccess choice, 10% lower if the project tries to turn from 0% collected to at least something and 0.1% lower chances when turning from collecting at least something to collecting all the sum. Overall, as we may see, females are less likely to collect at least something for their startups, but for those projects, which already collected a certain amount of money, the fact that the founder is female will influence only on less that 1% chance to not reaching the targeted sum. Thereafter, we may conclude, that for all other factors doing well (e. g. the targeted sum is not too high, the category chosen right, and the biography written long enough, it doesn't matter whether the project was made by male or female, it has almost equal chances to reach the targeted sum.

Next, we're turning to the question of whether it is relevant to make a website for a project and share your Facebook link on your Boomstarter project website. As a result of the first regression, we may see, that the projects with the website have the 14% higher probability to succeed. And, if the author adds his Facebook account, the probability to collect 100% of the targeted sum increases for 10%, which is reciprocally to the probability to collect at least something, shown by the second regression model.

Moreover, if we look at categorical variables, we may see that they are almost insignificant for model of chose between collecting 100% and at least something, which means that in this case of project category and city rage (by number of population) the most significant threshold of percentage gain is the one, which is from 0 to at least something as a percentage of targeted sum, cause it's the one where there is the highest increase in the probability of success.

As a result, we may see, that overall cross-country analysis showed, that overall trends, like the negative quadratic relation of the targeted sum to the probability of project success, and the positive linear trend for the number of followers are similar for all countries. However, when we turn to the categorical analysis, the relations are rather different. The adding of categories to the Russian crowdfunding platform regression model finished with negative coefficients, which means, that 5 most popular categories decrease the probability of the project success. Nevertheless, in case of the other countries, the categories are positive, which means that the projects in these categories have higher probability to succeed. As for the Boomstarter projects, the 8,000 RUB is a threshold maximum sum for the projects to get the 50% probability to collect 100% of the targeted sum for the project. However, this sum is only for the 50% confidence in obtaining the 100% funding, which increases to 400,000 RUB and more if there are 20% probability of success. As for the project categories on Boomstarter, there is an interesting fact, that categories and cities are significant only in case we do not include author-specific factors in our models. Further we'll explain which the main implications of this study for both entrepreneurs and researchers are.

4.1 Results discussion

The result of this study may be divided into two parts: the cross-country one, where there are main implications of cross-country regression models, and the one, mainly focused on Russian specific factors, with deeper research of the factors, influencing the project success on Russian crowdfunding platform. The first part may also be divided to the description of the factors, influencing the probability of getting 100% collection, rather than 0% and the one, describing the factors, influencing the probability to collect 100% rather than less than 100%, but more than 0%.

Firstly, we'll describe the factors, influencing turning the projects from unsuccessful to successful ones. As we've already mentioned in the results section, the maximum sum for the project, which we may advise to set for the projects, run in Australia, Canada, Great Britain or USA, is not higher than $6,000 in order to achieve the targeted amount with 50% probability. As for the Russia, the 50% probability of success in this case will be available for the projects with maximum targeted sum equal to 9,000 RUB. This implication fully approves our H1 about the inverse relationship between targeted sum and project success.

As for the number of investors, for all countries except Russia, overall trend is downward. However, if we take into account the 95% confidence interval, we may see that the number of investors remains the same, except for the bottom threshold, which goes to the lower success probability rate, when increasing the investors number. This means, that there are possible other factors influencing the investors number, due to the fact that we have the highest success probability with those projects with a low number of investors, and when the project is overwhelmed with the investors, this means that the project is rather really successful, or has a huge number of investors with small donations. In the second case we may also add, that such projects are usually unsuccessful.

Also, as we may see, the project categories influence of the project success also differs. At the very beginning of this research we pointed out that for our cross-country analysis we used only 5 categories, which are similar for both Boomstarter and Kickstarter. As a result, we may see that all these categories have significant influence on the project success. However, the Russian project have an intention to decrease their successfulness pertaining one of the categories, included in our regression models. Thereafter, as for other countries, we may see that these categories are positively related to the project successfulness. Overall, we may say, that by this we confirmed our 0 hypothesis about the differences in influential factors of project success on crowdfunding platform and also confirms the H2, which was about the influence of the category to the project success.

Next, we also understood, that as for Russia, there are also significant author-specific factors, influencing the project success. As we've seen, there is an optimal author's age for running a project. Moreover, we've also seen that the number of followers increases the probability of project success, as in case with 100% collection, so for the situation when the project tries to collect at least something, these verifies conclusions made by (Mollick, 2014) in case for the USA. As to the gender of the founder, we've seen that females have 20% less probability to collect the targeted amount. That confirms results of (Geiger & Oranburg, 2018) and (Koch, 2016) who stated that gender makes negative influence on collection rate in USA, but contradicts (Anglin et al., 2018) conclusions for a small UK crowdfunding platform. As to the biography length, it has a positive influence on the project success in Russia. Finally, as for the targeted sum for a project, as in other countries, it has a negative influence on the project success. It verifies conclusions made by (Forbesa & Schaefera, 2017) for the United Kindom, (Lagazio & Querci, 2018) for Italians on Indiegogo, (Mollick, 2014) for Kickstarter. However, there were only two out of 6 regression models for Russia, where the targeted sum logarithm was significant wh...