Search

Machine learning (ML) methods are proliferating in scientific research. However, the adoption of these methods has been accompanied by failures of validity, reproducibility, and generalizability. These failures can hinder scientific progress, lead to false consensus around invalid claims, and undermine the credibility of ML-based science. ML methods are often applied and fail in similar ways across disciplines. Motivated by this observation, our goal is to provide clear recommendations for conducting and reporting ML-based science. Drawing from an extensive review of past literature, we present the REFORMS checklist (recommendations for machine-learning-based science). It consists of 32 questions and a paired set of guidelines. REFORMS was developed on the basis of a consensus of 19 researchers across computer science, data science, mathematics, social sciences, and biomedical sciences. REFORMS can serve as a resource for researchers when designing and implementing a study, for referees when reviewing papers, and for journals when enforcing standards for transparency and reproducibility.

Any large dataset can be analyzed in a number of ways, and it is possible that the use of different analysis strategies will lead to different results and conclusions. One way to assess whether the results obtained depend on the analysis strategy chosen is to employ multiple analysts and leave each of them free to follow their own approach. Here, we present consensus-based guidance for conducting and reporting such multi-analyst studies, and we discuss how broader adoption of the multi-analyst approach has the potential to strengthen the robustness of results and conclusions obtained from analyses of datasets in basic and applied research.

How predictable are life trajectories? We investigated this question with a scientific mass collaboration using the common task method; 160 teams built predictive models for six life outcomes using data from the Fragile Families and Child Wellbeing Study, a high-quality birth cohort study. Despite using a rich dataset and applying machine-learning methods optimized for prediction, the best predictions were not very accurate and were only slightly better than those from a simple benchmark model. Within each outcome, prediction error was strongly associated with the family being predicted and weakly associated with the technique used to generate the prediction. Overall, these results suggest practical limits to the predictability of life outcomes in some settings and illustrate the value of mass collaborations in the social sciences.

Adult death rates are a critical indicator of population health and well-being. Wealthy countries have high-quality vital registration systems, but poor countries lack this infrastructure and must rely on estimates that are often problematic. In this article, we introduce the network survival method, a new approach for estimating adult death rates. We derive the precise conditions under which it produces consistent and unbiased estimates. Further, we develop an analytical framework for sensitivity analysis. To assess the performance of the network survival method in a realistic setting, we conducted a nationally representative survey experiment in Rwanda (n = 4,669). Network survival estimates were similar to estimates from other methods, even though the network survival estimates were made with substantially smaller samples and are based entirely on data from Rwanda, with no need for model life tables or pooling of data from other countries. Our analytic results demonstrate that the network survival method has attractive properties, and our empirical results show that this method can be used in countries where reliable estimates of adult death rates are sorely needed.

The network scale-up method enables researchers to estimate the size of hidden populations, such as drug injectors and sex workers, using sampled social network data. The basic scale-up estimator offers advantages over other size estimation techniques, but it depends on problematic modeling assumptions. We propose a new generalized scale-up estimator that can be used in settings with non-random social mixing and imperfect awareness about membership in the hidden population. Further, the new estimator can be used when data are collected via complex sample designs and from incomplete sampling frames. However, the generalized scale-up estimator also requires data from two samples: one from the frame population and one from the hidden population. In some situations these data from the hidden population can be collected by adding a small number of questions to already planned studies. For other situations, we develop interpretable adjustment factors that can be applied to the basic scale-up estimator. We conclude with practical recommendations for the design and analysis of future studies.

The network scale-up method is a promising technique that uses sampled social network data to estimate the sizes of epidemiologically important hidden populations, such as sex workers and people who inject illicit drugs. Although previous scale-up research has focused exclusively on networks of acquaintances, we show that the type of personal network about which survey respondents are asked to report is a potentially crucial parameter that researchers are free to vary. This generalization leads to a method that is more flexible and potentially more accurate. In 2011, we conducted a large, nationally representative survey experiment in Rwanda that randomized respondents to report about one of 2 different personal networks. Our results showed that asking respondents for less information can, somewhat surprisingly, produce more accurate size estimates. We also estimated the sizes of 4 key populations at risk for human immunodeficiency virus infection in Rwanda. Our estimates were higher than earlier estimates from Rwanda but lower than international benchmarks. Finally, in this article we develop a new sensitivity analysis framework and use it to assess the possible biases in our estimates. Our design can be customized and extended for other settings, enabling researchers to continue to improve the network scale-up method.

The network scale-up method is a promising technique that uses sampled social network data to estimate the sizes of epidemiologically important hidden populations, such as sex workers and people who inject illicit drugs. Although previous scale-up research has focused exclusively on networks of acquaintances, we show that the type of personal network about which survey respondents are asked to report is a potentially crucial parameter that researchers are free to vary. This generalization leads to a method that is more flexible and potentially more accurate. In 2011, we conducted a large, nationally representative survey experiment in Rwanda that randomized respondents to report about one of 2 different personal networks. Our results showed that asking respondents for less information can, somewhat surprisingly, produce more accurate size estimates. We also estimated the sizes of 4 key populations at risk for human immunodeficiency virus infection in Rwanda. Our estimates were higher than earlier estimates from Rwanda but lower than international benchmarks. Finally, in this article we develop a new sensitivity analysis framework and use it to assess the possible biases in our estimates. Our design can be customized and extended for other settings, enabling researchers to continue to improve the network scale-up method.

Individuals influence each others' decisions about cultural products such as songs, books, and movies; but to what extent can the perception of success become a "self-fulfilling prophecy"? We have explored this question experimentally by artificially inverting the true popularity of songs in an online "music market," in which 12,207 participants listened to and downloaded songs by unknown bands. We found that most songs experienced self-fulfilling prophecies, in which perceived--but initially false--popularity became real over time. We also found, however, that the inversion was not self-fulfilling for the market as a whole, in part because the very best songs recovered their popularity in the long run. Moreover, the distortion of market information reduced the correlation between appeal and popularity, and led to fewer overall downloads. These results, although partial and speculative, suggest a new approach to the study of cultural markets, and indicate the potential of web-based experiments to explore the social psychological origin of other macrosociological phenomena. (Contains 9 figures, 4 tables, and 11 footnotes.)