Background
Patriarchy's decline
Patriarchy was the judicially recognised norm for centuries in both Western and East Asian nations. Traditionally and legally, the father was entitled to the custody of the children of a marriage and could deny a mother access to her children. As Danaya Wright observed, ‘Judges in England in the eighteenth and most of the nineteenth centuries unhesitatingly enforced a father's unlimited right to custody of his children in every case where he had not forfeited his paternal rights through some malfeasance, such as cruelty or desertion.’ Footnote 1 Likewise, Michael Grossberg's assessment of judicial patriarchy in America concluded that ‘[t]he assumption of patriarchal authority by the bench rested to a significant degree on the nineteenth century's increasingly rigid segregation of homebound female and worldly male functions. It represented a refined version of the distinction between the male authority to govern the home and the female responsibility to maintain [it].’Footnote 2
Patriarchal approaches had declined, however, by the beginning of twentieth century and were gradually displaced by the ‘tender years doctrine’ favouring the mother in custody determinations.Footnote 3 Then responding to societal pressure for sexual equality and gender neutrality, in the 1970s and 1980s most US states eliminated the ‘tender years’ presumption, leaving a void filled with a ‘child's best interests’ standard to determine child custody.Footnote 4
In East Asia, countries such as South KoreaFootnote 5 and Taiwan,Footnote 6 influenced by Confucianism but experiencing a wave of democratisation in recent decades, faced enormous challenges in reconciling Confucian patriarchal ideology with the concept of gender equality found in their amended Civil Codes. As their societies democratised, both South KoreaFootnote 7 and TaiwanFootnote 8 started to shift from paternal privilege to ‘the best interests of the child’ by modifying statutes regarding child custody. However, the new gender-neutral standard may not in fact change the ‘law in action’ as soon as legislators expected.Footnote 9 What is the ‘law in action’ in Taiwan, and to what extent has it departed from the traditional Confucian paternal preference?
The legal framework regarding child custody in Taiwan
Taiwan's shift from paternal supremacy to gender-neutral rules regarding custody occurred relatively recently, in 1996. Previously, Taiwan's Civil Code had recognised numerous paternal preferences in parental rights, the assignment of children's surnames, and domicile and post-divorce arrangements. These patriarchal clauses became some of the major targets for legal reform over the past thirty years in Taiwan.
At first, reform of parental rights was part of the movement for women's rights and gender equality.Footnote 10 Up until 1996, Article 1089 of the Civil Code provided that in the event that parents disagreed on how to exercise parental rights over a minor, the father had the right of final decision. In 1994, in response to feminist movements in the Post-Martial Law Period,Footnote 11 Taiwan's Constitutional CourtFootnote 12 held Article 1089 to be unconstitutional in Judicial Yuan Interpretation No. 365.Footnote 13 In that pathbreaking decision, the Grand Justices of the Constitutional Court found that Article 1089 was inconsistent both with Article 7 of the Constitution, which proclaims that both sexes are equal under the law, and with Article 9, Paragraph 5 of the Additional Articles of the Constitution, which eliminates sexual discrimination. In their reasoning, the Grand Justices recognised the ‘best interests of the child’ standard, which had been adopted worldwide, for the first time.Footnote 14 This ruling laid the groundwork for the highly significant 1996 Civil Code parentage amendments.Footnote 15
Following Judicial Yuan Interpretation No. 365, the 1996 Civil Code amendments addressed not only Article 1089 but also Article 1055. The previous Article 1055 had stipulated that, in both consensual and judicial divorce, the custody of children belonged to the father unless either it had been agreed otherwise in a consensual divorce (Article 1051) or the court had decided otherwise (Article 1055). Prior to the amendment, there had been only one form of child custody: sole custody. The 1996 the Civil Code amendments repealed Article 1051 and amended Article 1055, replacing paternal preference with the ‘best interests of the child’ doctrine, and recognising joint custody and non-custodial parents’ visitation rights.
Child custody in reality
Both before and after the 1996 amendments, the vast majority of custody arrangements in Taiwan were made voluntarily between the parents rather than being judicially decided. For example, according to the Judicial Statistics Yearbook, for judicial divorces in 2020, the custody of 1,307 children were assigned.Footnote 16 By comparison, according to Ministry of the Interior statistics, the total number of children for whom custody arrangements were made in 2020 was 56,045.Footnote 17 So 98 per cent of child custody arrangements were made voluntarily between the parents.
As for the type of custody, information about custody arrangements prior to 2002 is unfortunately hard to come by, as no official statistics were published. Afterwards, official statistics demonstrate that, examining all divorce cases (both judicial and consensual) from 2002 to 2020, fathers had priority at first, although the gap between paternal and maternal sole custody gradually narrowed. Finally in 2018, the number of mothers granted sole custody exceeded the number of custody granted to fathers. In addition, joint custody has become more common (Figure 1).
Custody Arrangements in All Divorce Cases (Including Consensual and Judicial Divorce, 2002–2020)Footnote 18
Custody arrangements decided in judicial divorces show a picture different from non-judicial divorce cases. Prior to 1996, it is estimated that courts awarded custody to fathers in some eighty to ninety per cent of cases.Footnote 19 According to Hung-En Liu's empirical study analysing court decisions from 1998 to 2000 (the years immediately after the 1996 law reform), maternal sole custody was prevalent, although economic competence was regarded as a necessary, but not sufficient, factor in determining custody. This might have privileged fathers.Footnote 20 After 2002, sole custody continued to go chiefly to the mother (Figure 2). Contrary to a growing preference outside the judicial system for joint custody, the courts tended not to consider joint custody. This is because parents who are not able to agree to an amicable divorce and who enter the judicial process are believed to be highly conflicted, circumstances that courts tend to consider as unsuitable for joint custody.Footnote 21
The Assignment of Custody in Judicial Divorces (2002–2020)Footnote 22
Issues and literature review
To assist the court in determining the child's best interests, the new Article 1055-1 lists several factors that judges must consider, such as the age, sex, birth order, health condition, and the wishes of the child, and the age, occupation, character, economic ability and lifestyle of the parents, etc.Footnote 23 In fact, such a broad standard gives judges considerable discretion in deciding what is in the best interests of the child.Footnote 24
Although empirical legal studies methods have not been widely adopted by Taiwanese researchers,Footnote 25 including family law scholars, child custody is undoubtedly an exception. Several studies are available, all exploring the ‘law in action’ regarding custody.
Study on frequency of factors
Hung-En Liu's pathbreaking 2001 study collected seventy cases decided from 1998 to 2000 and categorised each of these cases by the factors that were considered by judges to decide the child's best interests. The results demonstrated that the courts tended to consider only some of the factors listed in Article 1055-1, such as ‘interview report of social workers’ (46 per cent), ‘occupation and economic resources of the parents’ (39 per cent), ‘wishes of the child’ (32 per cent), and ‘age of the child’ (27 per cent).Footnote 26 However, Liu's work did not discuss how these factors affected courts’ actual decisions.
Study on relationship between factors and outcomes
Similarly, Yen-Ni Cheng coded factors that were considered in each of 540 cases from 2012 to 2014. Additionally, she applied Pearson's chi-square testFootnote 27 to determine the relationship between each factor and the final custodian (father/mother sole custody, joint custody, split custody,Footnote 28 or third-party custody). Her finding was that ‘interview report of social workers,’ ‘wishes of the child,’ ‘primary caretaker,’ ‘current residence of the child,’ ‘wishes of the parents,’ ‘parent-child interaction,’ and ‘misbehavior of the parents’ were significantly associated with the choice of custodian.Footnote 29 However, in terms of the ‘age of the child,’ recognising that many cases mentioned the principle of ‘tender years’ but in contrast to Liu's conclusion, Cheng argued that child age is not significantly related to the choice of custodian. If the child was under two years old, Cheng found, 79.4 per cent of the cases awarded mothers sole custody. Likewise, when the child was between 12 and 17 years old, 80.6 per cent of the cases awarded mothers sole custody. Cheng concluded that no matter how old the child was, there was no difference in determining maternal preference in custody cases. The principle of ‘tender years’ was no more than a pretext for the courts to justify the result of maternal sole custody if the child was young.Footnote 30
Long-Term observation
Chao-Ju Chen conducted an empirical study collecting 272 cases decided between 2000 and 2013, making diachronic comparisons possible. This study characterised four doctrines – the ‘tender years’ doctrine, the primary caretaker doctrine, the principle of continuity, and the ‘friendly parent’ doctrine – as the four doctrinal pillars supporting the ‘best interests of the child’ principle and observed whether each of the four was applied in the cases Chen collected.Footnote 31 Chen observed that the significance of the ‘tender years’ doctrine has substantially diminished over the years. Chen also found that the courts tend to grant sole custody to mothers in cases where both parents seek custody, the majority of which involve ‘responsible mothers.’ Meanwhile, a father's chances of receiving sole custody granted by the courts are improving, especially when he is the only party who actively seeks custody, such as in the case of transnational marriages composed of a Taiwanese husband and a foreign wifeFootnote 32 who ‘ran away’ from Taiwan.Footnote 33 Therefore, Chen concluded that maternal preference is more a myth than a reality.Footnote 34
Through the examination of the texts of court decisions, the aforementioned literature points to several factors that have been frequently regarded by judges as important to the outcome of a case.Footnote 35 Nevertheless, a simple yet significant question has not yet been answered: among these factors, which ones take priority in Taiwanese judges’ minds? By relying on the remarkable development of computer algorithms, this question becomes possible to answer.
Prediction of court decisions
Automatic prediction of court decisions is an enormous field expanding drastically in the past few years.Footnote 36 Numerous studies have attempted legal prediction through machine learning technology,Footnote 37 often with impressive results. For example, Ruger et al compared two methods of forecasting US Supreme Court decisions: the first was a decision tree model which predicted 75 per cent of the Court's affirm/reverse results correctly, while the second was a set of predictions made by legal experts, getting only 59 per cent right.Footnote 38 Katz et alFootnote 39 used the data from the US Supreme Court database to forecast justices’ votes and reached an accuracy rate of 72 per cent through a ‘random forest’ classifier methodFootnote 40, while later Kaufman et al adopted a ‘boosted decision trees’ methodFootnote 41 achieving an accuracy rate of 75 per cent.Footnote 42 In addition to manually annotated data, some studies applied natural language processingFootnote 43 to deal with ‘unstructured’ raw judgment text data. Aletras et al used ‘support vector machine’ classifiersFootnote 44 to predict case outcomes of the European Court of Human Rights with an accuracy rate of 79 per cent on average.Footnote 45 As for appellate court cases, Raghupathi et al analysed cases on pharmaceutical patent validity in the United States Court of Appeals for the Federal Circuit by the Naïve-Bayes model,Footnote 46 achieving a 78 per cent overall accuracy rate.Footnote 47
Automatic outcome prediction methods have also been employed in analysing family court cases in various countries. A 1999 study adopted a ‘neural network’ approachFootnote 48 to analyse 400 Australian divorce cases to predict the assets that a husband or wife would receive upon property distribution after divorce.Footnote 49 In contrast with recent research applying natural language processing to unstructured judgment texts directly, this early study used raters to read the judgements and record values of variables. But the ‘neural network’ approach has difficulties in explaining outcomes, since the inferencing steps of this algorithm are not made explicit.Footnote 50
In 2021, Muñoz Soro and Serrano-Cinca applied neural network and logistic regression to 1,884 Spanish appeal court rulings (second instance) on child custody and predicted whether the court would grant sole custody or joint custody with an accuracy exceeding 85 per cent.Footnote 51 They found that ‘“the relationship and attitudes of the parents” and “the psychophysical circumstances of the child”’ are the two factual elements that the judge took most into account when deciding the type of custody.Footnote 52 However, this study did not provide insights on what the contents of the best interests of the child are.Footnote 53 And the second instance appeal judgments may have selection bias issues, such as the effects of settlement of cases pending appeal.Footnote 54
Regarding Taiwan, in a 2017 study we employed a neural network approach to predict court decisions on child custody in 2012–2014 with an accuracy rate of 98.6 per cent,Footnote 55 but we were not able to identify factual elements that influenced court decisions. That result is not sufficiently useful for legal professionals. Compared to the 2017 article, this Article incorporates more recent data for 2015 through 2017 and uses a different tool. Our goals encompass not merely the prediction of custody outcomes, but also the identification of the factors that judges consider first in custody cases when ascertaining the ‘best interests of the child.’ To accomplish these goals, we established a dataset and carefully coded it to determine the specific factors considered. Instead of a ‘neural network’ approach, we adopted decision tree learning in order to attain a more explainable result, as discussed below.
Research design
This section elaborates on how the data sample was collected and coded, and how decision tree learning was applied.
The data
All cases except juvenile and sexual assault cases decided by district courts in Taiwan since 2000 are open to the public on the official website of the Judicial Yuan.Footnote 56 We focus on child custody decisions decided by district courts at first instance because the vast majority of child custody cases are handled by these courts, few of their decisions are appealed, and almost none of the appeals succeed.Footnote 57 Using carefully chosen causes of action, keywords and decision dates,Footnote 58 we identified 3,028 child custody decisions between 1 January 2012 and 31 December 2017.Footnote 59
We then limited our sample to cases in which both parents were Taiwanese and both sought to acquire custody. This is because sometimes one parent (usually the defendant) does not come to court or keeps silent, as frequently occurs in transnational marriage cases. In these cases, social workers have difficulty conducting interviews, and judges may not receive sufficient information about absent or silent parents. In cases of default judgement, plaintiffs are very likely to receive custody.Footnote 60 Judges do not need to weigh the factors between two parties to decide custody – it is almost automatically granted to the party who showed up (the plaintiff). So, we excluded these cases from our dataset. Among the 3,028 cases, the 2,096 cases in which one of the parents did not express any opinion regarding the choice of custodian and the 97 cases of transnational marriage were therefore excluded.Footnote 61 The remaining 835 cases (3028-2096-97 = 835) contain 1,290 children. Among them, 1,126 children were placed under sole custody (87.3 per cent), 159 under joint custody (12.3 per cent), and five under third party custody (0.4 per cent) (Figure 3).
Types and Percentage of Custody Arrangements in Taiwan's District Courts (2012–2017)
As pointed out in previous literature, parents’ willingness to enter joint custody is one of the most relevant factors for the court to grant joint custody.Footnote 62 Since joint custody is more desirable due to its better outcomes than sole custody,Footnote 63 if both parents are willing to enter joint custody the court usually follows parents’ wishes without comparing each parent's performance in such detail as it does in sole custody cases. That is to say, the court first decides whether joint custody is a possible option. If it is not, then the court weighs each parent's performance and selects one parent as sole custodian.Footnote 64 Our goal here is to capture courts’ decision-making process in this last step: comparing parents’ performance. Therefore, we include only the 1,126 children in the sole custody cases in our final sample.
Dataset construction
Court judgements in Taiwan are publicly available but only in raw text style, ie, they are unstructured narrative data. Moreover, in divorce judgments, parties’ names are often redacted for privacy reasons. Therefore, coding variables from raw texts is a difficult process.
Judgments of Taiwanese courts typically contain four sections: the title (including the case number, the parties, and their lawyers); operative provisions (the case outcome, whether the divorce was upheld, and if so, the child's custodian),Footnote 65 facts and reasons (including parties’ arguments, the facts recognised and reasons articulated by the court), and miscellaneous items (the judgement date and the names of the judge and clerk).
The custodian outcome of the case, ie, whether the mother or the father receives custody, is our dependent variable. However, this outcome is not easily extracted automatically by machine from the operative provisions or the facts and reasons. Parties’ names are often anonymised, so it is not evident whether the plaintiff is the mother or the father. But close attention to the operative provisions and the facts and reasons – a task human researchers can best perform – can often resolve this point.
Take, for example, the case Taipei District Court 102 (2013) Hun No 183. Footnote 66 The published opinion confirms that the plaintiff won the divorce litigation and was granted custody. But names of the parties are anonymised, so the plaintiff's gender is unclear. However, the report of the guardian ad litem mentions that the plaintiff is breastfeeding the child. So, we know that the plaintiff was the mother.
Other cases illustrate the point. Phrases such as ‘defendant offers the child sufficient father's love (父愛) and care,’Footnote 67 or ‘plaintiff is the primary caretaker who provides love and care so that mother and daughter have good conversations on life and education (母女二人於生活及教育上有共同話題)’Footnote 68 appear often in the section on facts and reasons. Human researchers can determine the parties’ gender in such cases more easily than machines can. So, we coded case outcomes manually.
To ascertain the independent variables, after examining the provisions of Article 1055-1, social workers’ evaluation items and existing literature, we identified nineteen factorsFootnote 69 that judges may consider in custody cases (Table 1).
Nineteen variables (factors) that Family Court judges may consider in child custody decisions
Coding is based on each individual child, as multiple children in the same family may have different ages, sexes, wishes, interactions with parents, etc. The coding rule is simple: due to a child's right to live with his or her parents (as stipulated in the Convention on the Rights of the Child),Footnote 70 if it is not against the best interests of the child, the judge should grant custody to a parent. Therefore, if both parents meet the minimum criteria to be custodian, then the judge selects the parent considered more capable (since joint custody is not often preferred).Footnote 71
To capture this decision-making process statistically, we adopt the following approach. We take the judge's conclusion regarding the relative strength of the mother's and the father's performance on each of the nineteen factors and we score it.Footnote 72 As for scoring of independent variables, for example, for the variable ‘parenting time’, if the judge concluded that the mother performs better, we categorise it as 3. If the conclusion was that the father performs better, we label it as 1. If the performance is equivalent (ie, both mother and father perform well or badly) or not mentioned by the judge, the category is 2. Thus, our variables are all categorical.
Decision tree learning
Decision tree learning is a method commonly used in data mining and machine learning.Footnote 73 The reason to adopt this approach is that its predictive power performs better than statistical techniques such as regressions.Footnote 74 Decision trees have ‘the advantage over logistic regression of … not depending on underlying assumptions about the distribution of the explanatory variables.’Footnote 75 In addition, decision tree procedure inherently reveals key interactions among all predictor variables without the need for the analyst to specify them a priori. It produces an output that conforms to the hierarchical and dichotomous nature of judicial decision-making.Footnote 76 And decision trees are easily interpretable.Footnote 77 Since our goal is to create a model that predicts the value of a target variable (here, custody granted to father or mother, 1 or 0, a dichotomous structure) based on several input variables (here, the 19 factors listed above), and to identify top factors, variables, and predictors, we chose decision tree analysis as a tool.
A decision tree is a flow-chart-like structure, where each internal (non-leaf) node represents a feature, and each branch represents a value the node can assume. During the building process, input data are split according to the evaluation of selected features (denoted as variables here) starting at the root node, then one at a time until the subset of a node shares the same target variable or splitting does not influence the target variable. After obtaining the tree, fed instances are classified starting at the root node and sorted based on their feature values, to determine the target variable result. This study adopts the CHAID (Chi-squared Automatic Interaction Detector) algorithm to make strategic splits.Footnote 78
As is usual for machine learning experiments, we ran the decision tree learning process through random sub-sampling validation in the following steps: (1) 226 (20 per cent) of the children were randomly designated as test cases and taken out of the data sample, hiding their outcomes. (2) The remaining 900 children (80 per cent) were used as a training set for the algorithm. (3) The trained algorithm (model) predicted the outcome of the test cases. (4) The predicted outcome was compared to the test cases’ previously hidden real outcomes to see if the predictions were correct, mistaken, or absent, and the result was recorded. (5) The preceding steps were repeated 1,000 times. (6) The results of the 1,000 trials were then averaged, in order to derive the following accuracy rates. In this way, the bias is smaller than randomly partitioning the training set and the test set and only performing training and testing once.
Results
Model efficacy
In supervised training, to evaluate the performance of a machine-built classification model, a confusion matrix is usually used to visualise the results of the test set.Footnote 79 The matrix is composed of two dimensions, with the data column containing results predicted by the machine and the data row containing the actual results (see Table 2).
Confusion Matrix of Child Custody Test Set
Seed = 7739
A set of four indices is commonly used to evaluate model performance during the machine learning process. The first and the most instinctive index is accuracy, the percentage of correct predictions by the machine, defined as (TP+TN) / total sample size. The second is precision, the proportion of true positives among all samples classified as positive, defined as TP/(TP+FP). The third index is the rate of correctly classified true positives, named recall, defined as TP/(TP+FN). And fourth, to take both precision (P) and recall (R) rate into account in case of a conflict, the F1 Score (the ‘F-measure’) is used to evaluate the model. The F1 Score is the harmonic mean of the two indices, defined as F1 score = 2/(1/P+1/R).
Our model's accuracy is 96.5 per cent in its test set and its F1 score is 0.9783, indicating that the model is quite satisfactory.
Model demonstration
For the custodian outcome of the case, ‘1’ represents sole paternal custody; ‘0’ represents sole maternal custody. As shown in Figure 4, the left number ‘.75’ under the root node indicates that 75% of test set cases are classified ‘0’ (custody for mother) and right number ‘.25’ means 25 per cent of the cases are classified ‘1’ (custody for father). The feature/ independent variable under the root node is ‘caregiver’ (marked in the branch). This feature best divides the data, which indicates that judges in Taiwan primarily consider who the current caregiver of the child is in the first place. Regarding the independent variable ‘caregiver,’ if the judge considers that the mother performs better (the mother is the primary caregiver), we categorise it as 3. If the judge concludes that the father is the primary caregiver, the label is 1. If the performance is equivalent (ie, both mother and father are caregivers) or not mentioned by the judge, the category is 2. Thus, in Figure 4, the black line (branch) on the left side shows that when the caregiver ≥ 2, ie, if the caregiver is the mother (represented by the number ‘3’) or both mother and father (represented by the number ‘2’), it comes to the next node. Here, 92 per cent of cases resulted in maternal custody and 8% in paternal custody. Also, the number ‘78%’ in this node means that ‘caregiver ≥ 2’ shares 78 per cent of the whole sample, ie, 176 children (226 x 78%), while ‘caregiver < 2’ (father is the primary caregiver) shares the remaining 22 per cent (226 x 22% = 50 children).
Decision Tree of Child Custody Cases in Taiwan
Next, the variable ‘childWill’ appears, meaning that the model predicts that judges will, secondarily, consider the child's wishes. If the child prefers the mother or has no preference (≥ 2), the model again follows the black line (branch) on the left side coming to the end node, which is labelled with a probability distribution indicating that in the test set, custody has a very high probability (95 per cent) of going to mother (the label ‘0’). On the contrary, even if the mother is the primary caregiver, in the case that the child prefers the father (‘childWill < 2’), the probability that the mother gets custody becomes fairly low (17 per cent in the test set). It can be observed that on the right branch, representing when the father is the caregiver (‘caregiver < 2’, meaning that caregiver = 1) of the root node, judges next observe the interaction between parents and the child. If the father performs as well as or better than the mother (‘interaction < 3’, meaning that interaction = 1 or 2), we see ‘.90’ at the right end node, meaning that the father has a high probability (90 per cent) of being the custodian. In contrast, if the mother interacts with the child better than the father (‘interaction ≥ 3’), the mother is likely to acquire custody (79 per cent).
Discussion
Previous literature attempted to determine what factors affect custody decisions by surveying judges or by analysing court cases with statistical tools. This study, by contrast, used decision tree learning, an artificial intelligence technique, to answer a very specific question: what factors affect cases resulting in sole custody when both parents actively seek custody? With the assistance of cognitive computing and artificial intelligence, our research method can offer more detailed, practical information and a much more comprehensive picture of custody decisions in Taiwan.
We demonstrate that among the numerous factors stipulated in Article 1055-1 of the Taiwan Civil Code, ‘primary caregiver,’ ‘child's wishes’ and ‘parent-child interaction’ are the three most significant factors contemplated by judges. This pattern of judges’ decision-making regarding child custody appears to be relatively constant and stable during the six-year period studied.Footnote 80 Although ‘occupation and economic resources of the parents’ was once considered an important factor in child custodyFootnote 81 and is still widely accepted by the public,Footnote 82 our analysis does not identify it as significantly affecting judges’ decisions nowadays.
Furthermore, in custody disputes addressed by judicial decisions, the mother seems to have overwhelming supremacy: in our dataset, mothers had a 75 per cent likelihood of receiving sole custody (see the root node in Figure 4). But this cannot simply be attributed to judges’ preference for mothers over fathers. The majority of Taiwanese cases are those in which the mother happened to be the primary caregiver and the child prefers the mother, represented by the leftmost end node in Figure 4. Meanwhile, as previously mentioned, the most significant factor judges consider is ‘primary caregiver,’ suggesting that the reason that mothers are given custody more readily than fathers is not because of their gender, but simply because in most cases they are already the primary caregivers. This supplements Chen's observations that Taiwanese courts tend to grant sole custody to mothers in cases where both parents seek custody. Chen attributed this to ‘responsible mothers’.Footnote 83 Our study finds that when a father is the primary caretaker, ie, a ‘responsible father’, although fewer in number (only 22 per cent of our sample), he also has high chances (85 per cent) to acquire sole custody (the node in the second row in Figure 4). In this sense, we agree with Chen that maternal preference is a myth.
Another concern arises: if the mother loses the status of primary caretaker due to, for example, fleeing the marital home, will the mother fail to acquire child custody? As seen above in Figure 4, when the father is the caregiver, judges in Taiwan look into the interaction between parents and the child. If such a mother interacts with the child better than the father does (these cases are rare, only 2 per cent of our whole sample), the mother still has a high probability of being granted custody (79 per cent). This means that a mother fleeing from home may still have opportunities to be appointed custodian unless she was completely absent from the legal proceedings.Footnote 84
Conclusion
This study, aiming at understanding the ‘law in action’ in Taiwanese family law, clarifies how Taiwanese judges actually apply their statutory obligation to decide custody battles in the ‘best interests of the child.’ Our study not only identifies the outcome of each case (typically, sole custody going to mother or father), but also pinpoints the three key factors that judges contemplate in making their decisions in accordance with the legal standard. Those key factors underlying the outcomes are: (1) ‘primary caregiver’, (2) ‘child's wishes’, and (3) ‘parent-child interaction’.
This research should also help legal scholars identify particular custody cases as either typical or exceptional by comparing the machine-predicted results to actual judgements. A case may be exceptional, worth further exploration and research, if the two outcomes are inconsistent. In addition, this study should assist parents and their lawyers to preliminarily evaluate their possibilities of acquiring custody. Family lawyers can save time and costs in giving legal advice. And when the outcome of litigation can be predicted in advance by the parties, the likelihood of cases going to court will fall and the likelihood of settlement will increase. Thus, machine learning helps us predict what the ‘law in action’ is and hence contributes to legal certainty.
A limitation of this study is that the features we input into the model are based on the facts recognised by the judge, not the mother's or father's version of the facts. If the parties dispute the facts and hence one cannot be sure of these features, our model may not be able to offer accurate advice. Nevertheless, our research can help the parties narrow the disputes to facts and factors which are most likely to be considered by the judge, rather than litigating all potential issues and factors listed in the statutes. Similar methods might offer useful sources of predictive information to parents and their attorneys and to scholars in other countries in which patriarchy is in decline. In sum, our results should assist parents’ legal advisors, speed custody determination arrangements, and pave the road for future studies.
Open access
