Differentiating between statistical significance and substantive importance

1. Differentiating between statistical significance and substantive importance Jane E. Miller, PhD

2. Overview • Substantive significance defined • Quick review of statistics • What questions can they answer? • What questions can’t they answer? • How to implement a balanced presentation of multivariate results. Both • Statistical significance • Substantive importance

3. Objective of most research papers • Few people who write about multivariate analysis are focused solely on statistical mechanics such as developing new computer algorithms or formal statistical tests. • Some statisticians and methodologists will have those interests. • Most of us are interested in studying some relationship among social science or health concepts. • Test a hypothesis, derived from theory or previous empirical studies. • Inferential statistics are a necessary tool for hypothesis testing in quantitative research.

4. What is substantive significance? • Substantive significance of an association between two variables. • “So what?” • “How much does it matter?” • Real-world relevance to topic • In various disciplines, substantive significance = • “clinically… • “economically… • “educationally… • …meaningful” variation.

5. Example: Body Mass Index & mortality • Body mass index (BMI) shows a statistically significant positive association with mortality. • But is that gradient substantively significant? • Is it worth designing an intervention to decrease BMI as a way of decreasing mortality?

6. Key criteria for assessing substantive significance • Is the association causal? • Will changing the hypothesized cause lead to change in the purported effect? • Will weight loss (reduced BMI) yield lower mortality? • Is the effect big enough to matter? • Is the excess mortality among overweight or obese persons large enough to justify a program? • Can the hypothesized cause be changed? • Is BMI malleable?

7. Example prose • “For every hour a boy played a video game, he read just two minutes less than a boy who didn’t play video games. Notably, non-gaming boys didn’t read much at all either, spending only eight minutes a day with a book.” • From a NYT summary of Cummings and Vandewater, 2007.“Relation of Adolescent Video Game Play to Time Spent in Other Activities,” Archives of Pediatrics and Adolescent Medicine.

8. Quick review of statistical significance testing

9. Start with a hypothesis • The authors hypothesized that the more time adolescents spent on video games, the less time they spent on homework. • So far, description is purely in terms of the concepts under study. • No statistical jargon, yet… • To formalize this for statistical testing • Homework time = dependent variable (Y) • Gaming time = independent variable (Xi) • Ha= gaming time is negatively associated with homework time. • In other words, Xi is inversely associated with Y

10. Contrast it against the null hypothesis • The assumption of “no difference between groups” is called the null hypothesis (H0). • In the study on effects of gaming on homework time H0: time among gamers = time among non-gamers OR time among gamers - time among non-gamers = 0 • In words, the null hypothesis states that there is no difference in the amount of time spent on homework by gamers versus non-gamers.

11. What ? does inferential statistics answer? • “How likely would it be to obtain a difference at least as large as that observed between groups in the sample if in fact there is no difference between groups in the population?” • The p-value tells us the probability of falsely rejecting the null hypothesis. • Conventional levels of “statistical significance” : p<.05 • Strictly speaking, p<.05 tells us that for a large sample such as that used in the gaming study (N~1,400), the estimated effect size on time spent gaming is at least 1.96 times its standard error.

12. What questions DOESN’T it answer? • Whether the relationship is • Causal • Association ≠ causation • In the expected direction • The difference could be statistically significant but in the opposite of the hypothesized direction. • Big enough to matter in the real-world context • Each hour spent gaming reduced reading time by 2 minutes. Is that enough to induce genuine concern from parents or teachers? • Malleable

13. Conclusion: Don’t stop at “p<.05”! • “p<.05” answers only part of what we want to know about our research question. • It is a necessary but not sufficient part of statistical analysis. • Also need to consider questions about: • Substantive significance • Direction • Size • Causality • Non-causal associations should not be used to inform policy or program changes. • Confounding or spurious associations should be ruled out. • Often why a multivariate model is estimated.

14. Substantive significance overlooked • Many statistics textbooks show how to assess and present statistical significance. • Few if any show how to assess and present substantive significance.

15. Balance presentation of statistical and substantive significance • How to include both: • Inferential statistics for formal hypothesis testing. • Interpretation of substantive significance of findings in the context of the specific research question. • Critical for policy-makers and others not formally trained in statistics.

16. Principles for presenting results • Name the specific variables. Avoid • Writing about “my dependent variable” or “the effect size.” • Using acronyms from your database  • Report numbers in tables. • Interpretnumbers in text. • Incorporate units and categories for variables into the prose description.

17. What to report when comparing numbers • Direction (AKA “sign”) • For categorical independent variables (IV), which category has higher value of the dependent variable (DV)? • For continuous IVs, is the trend in the DV up, down, or level? • Magnitude • How big is the difference in the DV across values of the IV? • Statistical significance

18. Gender as a predictor of birth weight • Poor: “Boys weigh significantly more at birth than girls.” • Concepts and direction but not magnitude. • Statistical significance is ambiguous: Is the term “significant” intended in the statisticalsense or to describe a large difference? • Slightly better: “Gender is associated with a difference of 116.1 grams in birth weight (p<.01).” • Concepts, magnitude, and statistical significance but not direction: Was birth weight higher for boys or for girls? • Best: “At birth, boys weigh on average 116 grams more than girls (p<.01).” • Concepts, reference category, direction, magnitude, and statistical significance.

19. Substantive significance in the discussion • Place findings back in the broader perspective of the original research question. • Do they correspond to your hypothesis in terms of • Direction (sign) of the effect? • Size? • Whether the effect size was attenuated when potential confounders or mediators were taken into account? • What is the evidence for a causal relationship? • If not causal, what explains the association? • If causal, what are the implications for policy, programs, etc.?

20. A substantive issue from gaming study • “But the meaning of the finding [that girls who are gamers spend less time than non-gamers on homework] is not clear, as high-academic achievers often spend less time on homework as well.” • Places the finding in broader context by discussing other correlates of homework time.

21. Another substantive issue • “Although only a small % of girls played video games, our findings suggest that gaming may have different social implications for boys than for girls.” • Raises the question of selection effects: which girls play video games, and do their other characteristics affect how they spend their time?

22. Relate findings to previous studies’ • Are your findings consistent with the published literature on the subject in terms of statistical significance, sign, and approximate size? • If not, why not? • Different sample (place, time, subgroup) • Different data source or study design • Different model specification • Included potential confounders not previously analyzed. • Tested for possible mediating effects of 1+ factors.

23. Statistical significance in the discussion • Describe in words, not numbers. • No detailed standard errors, p-values, or test statistics in the discussion section. • Focus on the purpose of the statistical tests • Did the main variable of interest increase proportion of variance explained by the model? • Did some other variable “explain” the association between your key variable and the outcome?

24. Summary • Emphasize the substantive issues behind the statistical analyses. • Design the specification to match topic and data. • Choose plausible, relevant numeric contrasts. • Aim for a balanced presentation of statistical significance and substantive importance. • Use prose to ask and answer research question. • Use tables to report comprehensive, detailed statistics. • Use charts if needed to convey complex patterns.

25. Suggested resources • Chapter 3 (Statistical significance, substantive significance, and causality) in • Miller, J.E., 2015. The Chicago Guide to Writing about Numbers, 2nd Edition • Miller J.E. and Y.V. Rodgers, 2008. “Economic Importance and Statistical Significance: Guidelines for Communicating Empirical Research.” Feminist Economics. 14(2):117-149.

26. Suggested online resources • Podcast on comparing two numbers or series

27. Suggested practice exercises • Study guide to The Chicago Guide to Writing about Numbers, 2nd Edition. • Questions #2 and #4 from the problem set for chapter 3 • Suggested course extensions for chapter 3 • “Reviewing” exercises #1–3 • “Writing and revising” exercises #1–3

28. Contact information Jane E. Miller, PhD jmiller@ifh.rutgers.edu Online materials available at http://press.uchicago.edu/books/miller/numbers/index.html