1 Tonight ? Causal Inference and Internal Validity ? Sampling HD FS 503 Research Methods in Human Development & Family Studies Susan Hegland January 28, 2002 2 3 The Research Chain Links to previous studies Explanation, rationale, theory, point of view Questions, hypotheses, models Design Chapter 7 Participants Situations Causal Inferences and Internal Validity IV Observation DV Procedure Analysis & results Conclusion Links to next study 4 Internal Validity (Linking Power) ? Explanation credibility: Internal Validity: Hoffman-Reim • previous research? critical variables ? causal relationships in hypotheses? ? Translation fidelity Apply criteria for internal validity to the Hoffman-Reim study: Has the author provided sufficient support to ensure: ? Explanation credibility? ? Translation fidelity? ? Demonstrated result? ? Rival explanations eliminated? ? Credible Results? • Participants? • Instruments (test reliability & validity?) • Procedures ? Demonstrated result • Statistics appropriate? • Percentage of variance convincing? ? Rival explanations eliminated ? Credible result ? What about Zimbardo et al.? Hoffman-Reim? • Controls for alternate explanations? • Was hypothesis supported? 5 6 1 A Source of Potential Confusion Complexities in the Concept of Cause ? Internal Validity • Linking power within the study components ? Which is the cause in the causal chain? ? External Validity ? Causes as inferences • Generalizing power to the population ? Validated causal propositions escape disconfirmation ? NEVER confirmed! ? Test Validity (see chapter 18) • • • • Measurement reliability and validity Does the test produce consistent results? Does the test measure the construct? Does the test predict what it is supposed to? 7 Does laziness lead to school failure? Necessary & Sufficient Conditions Is the cause sufficient to produce the effect? Is the cause necessary to produce the effect? 8 Yes No Yes The one and cause!!! A contingent condition No An alternative condition A contributing condition ? Is laziness the one and only one cause of school failure? ? Is laziness an alternative condition? ? Is laziness a contributing condition? ? What’s the direction of causality here? ? Does school failure lead to laziness? ? Is there a causal chain here? 9 10 Drug abuse Contingent causes? ? Availability of drugs in community ? Prior use Contributing causes? ? Poverty ? Peer use ? Boredom Alternative causes? ? Prior addiction Simple patterns ?A ? B ? Infant attachment patterns cause adult attachment patterns ? Sensitive, accepting, cooperative, accessible parenting causes secure infant attachment ? Rarely appropriate! ? We live in a multivariate world 11 12 2 Multiple Causes Multiple Causes Positive Preschool Experience Family Environment Personality Work Kindergarten Achievement Marital Relations Parenting Social Network Family Resources Child Characteristics Ramey & Ramey, 1997 Belsky, 1984 13 Causal Chain? Positive parent interactions More family interactions Positive mother-child communication 14 Support for causal explanation Causal explanations are hard to establish in social sciences; they require: ? Precedence of cause in time ? Presence of effect does not reflect a chance deviation ? Explanation and evidence are congruent ? Effect follows a detailed prediction ? Effect follows the pattern of a manipulated cause ? Rival explanations eliminated Note: study this material carefully—it WILL be on the exam! Positive parentChild interactions Shared family views Less teen sex 15 Triangulation: establishing multiple supports for conclusions 16 The heart of description and explanation is the relationships between ? attributes (values) and ? variables Landmark paper: Baron & Kenny, 1986: ? Many researchers confuse ? moderating and ? mediating relationships ? Different roles, different statistical tests! ? Multiple sightings from multiple angles ? Multiple-measure, multiple-method procedures ? Multiple measures • Self-esteem and depression measures ? Multiple method • Parents and teachers rating social skills • Self-report and observers rating competence 17 18 3 Complex views of the world Moderating relationships ? Preexisting condition ? Affects when X causes Y ? Simple patterns are rare ? Multiple cause patterns are more common • Under what circumstances • Moderator variables establish when, or under what conditions • Mediator variables establish how: what’s the mechanism of influence • Baron & Kenny, 1986 ? Examples of moderating variables Sex Age (developmental level) Marital status Financial resources Culture Marital status Coping skills Social support network ? Tested as an interaction (i.e., sex x treatment) 19 Showing a moderating relationship Mother’s responsiveness 20 Another complicated study Child’s attachment Parent educational level Child reading skills Quality of marital relationship 21 What’s the cause here? 22 A mediating relationship ? Is parent education the cause? ? Or does some other variable mediate the relationship between parent education and children’s reading? Parent educational level 23 Child reading skills 24 4 A mediating relationship Mediating relationships ? Describe how cause affects outcome Parent-child language interactions ? When added to the model, ? “cause” is no longer related to the outcome Child reading skills Parent educational level ? tested in a hierarchical regression: see if mediator (i.e., language) accounts for some of the variance originally accounted for the the predictor (i.e., parent education) Baron & Kenny 1986 25 26 Is school size the cause? Register editorial School size and academic achievement ? Potential Moderators? Mediators? ? Moderators Mean ACT Scores 23 • Parent education? • Parent income? 22 21 ? Mediators 20 19 • • • • 18 17 < 250 250-399 400-599 600-999 10002499 25007499 7500+ Enrollment Quality of teachers (i.e., salaries)? Number of advanced courses? Ratio of extracurricular activities? Percentage of kids who take the ACT? ? How could you test? 27 28 IV and DV may be related but You won’t find it because ? Too much error ? Too small a sample ? Inappropriate measures ? Wrong sample ? Confounding variables ? Too stringent statistical criterion (alpha level): Type II error Mediators and Moderators Day care and aggression ? Possible mediators? ? Possible moderators? Family therapy and quality of marital communication ? Possible mediators? ? Possible moderators? Your studies? 29 30 5 32 IV and DV may be unrelated but You conclude they are related because ? treatment effects confounded with systematic error (e.g., response set, social desirability) ? confounding variables ? sampling error Too loose a statistical criterion ? (alpha level) ? Type I error Sampling 31 Teen sex prevention program and teen sexuality prevalence Sampling and Representation ? 300 high school students, how would you Sampling procedures ? Strategies for selecting a small number of units from a population ? Enable researchers to make reliable inferences about the population choose them? ? What steps would you go through to select them? ? Why? 33 Depression in freshmen: incidence and relationship to stress 34 Required sample size is based on: ? Certainty of inference desired • What’s the cost of drawing the wrong conclusion? ? Precision of inference desired • Population estimates must be very accurate • Or base N on prior studies (significant results!) ? Homogeneity of population on characteristic ? Goal: 200 first-year female students ? Random sample: N = 133 • Stratified by parents’ education and race • Representative of freshman class • multiple cultures? • many differences in parenting behaviors! ? Add a convenience sample • 67 cooperative females from dormitory ? ? Are you comfortable throwing these Size of effect vs normal sampling variation • small effect and large sampling variation ? samples together? Why or why not? very large study 35 36 6 Sample Size and Statistical Significance ? Treatment effect: big or small? Error: big or small? Statistical tests like t and F compare difference due to treatment (IV) difference due to error ? Treatment: potential cause controlled by IV Error: 25 Number of Participants ? • Variation in DV NOT explained by IV • Residual • noise ? More error means bigger (treatment) effect is needed (more treatment effect; more powerful treatment) or bigger sample is needed ? 20 15 Treatment Control 10 5 0 0 10 20 30 40 50 60 70 80 90 100 Scores In your review, include sample size! 37 38 To increase likelihood of finding a significant effect….. Treatment effect: big or small? Error: big or small? ? Maximize group differences due to treatment ? Minimize group differences due to error • Intensify treatment 50 Number of participants 45 Treatment Control 40 35 • Choose more homogeneous sample • Eliminate potential confounds 30 25 • Exclude some potential moderating variables 20 ? 15 Example: Zimbardo: 10 • only males 5 • only those susceptible to hypnosis • eliminate sources of systematic error (e.g., response sets, practice effects) 0 0 10 20 30 40 50 60 70 80 90 100 Scores 39 Sampling 40 Validity (accuracy) in sampling ? External validity: ? The process of selecting the people you want to study ? Want your sample to represent a larger population ? Only an issue if you hope to generalize your findings! ? In general, use probability, random sampling 41 (Generalizing Power) systematically (and randomly) sample from population to which you hope to generalize findings ? Internal validity: (Linking Power) randomly assign participants to treatments 42 7 Representative samples ? All members of the population have an equal chance of being selected • protects against bias • allows generalizability of findings ? Each sample member is independent, or unrelated, to the others ? Similarity among members in one treatment is due to treatment • Similarity is not due to other factors • Systematic error in sample looks like treatment (or independent variable) effects! Study of depression in freshman students ? Go to dorms and find who’s there at 8 PM on Saturday night in November ? First 100 students interviewed ? Conclusion: Nearly 50% of freshman are depressed; major intervention is needed!! ? Are you convinced? ? How should the sample have been drawn? 43 44 Independently select each case Probability Sampling ? Cannot measure the same person twice • As if the data came from two persons!!!!!!! ? Stratified Sampling ? Always check N ? Proportional Stratified Sampling • cannot be higher than number of cases! ? Oversampling ? Two recent dissertations ? Systematic sampling • N = 30, observed 9 times; n = 270 • N = 40, observed 8 times; n = 320 • Use a repeated measures statistic!!!! ? Cluster sampling ? Unit of analysis is child, not child at time x! 45 If I choose one husband to be a participant in my study, 46 Statistical tests look at N ? Smaller sample requires bigger treatment effect for the same treatment difference to be significant ? Bigger sample requires smaller treatment effect for the same treatment difference to be significant ? All statistical tests assume each case independent of other cases! ? Is the selection of a wife independent? ? If I choose a child… ? Is the selection of parent independent? ? Therefore, what is the case? ? The individual participant? ? Or the dyad? 47 48 8 Simple Random Sampling Types of Sampling ? Each unit of the population has an equal chance of being selected ? Choice of any unit is independent of choice of others ? Sampling frame = list of units from which we draw sample • Faculty directory • School directory • Telephone directory • List of licensed child care centers in county ? Probability Sampling ? Nonprobability Sampling 49 50 Probability Sampling Nonprobability sampling ? Requires random sampling of units from ? Selection of measures, treatment, etc. population ? Allows generalization of findings to population ? Selects participants or situations for study ? Sometimes used for convenience ? The rarer the population • the more expensive the sample • reason for nonprobability sampling • e.g., clinical samples 51 Sampling unit ? The family? child? classroom? student? ? Sampling unit is smallest unit receiving treatment • If teacher-taught curriculum, • the classroom is the unit, not the child!! • If computer-based curriculum • the child is the unit • If parent-child interaction • the dyad is the unit ? ANOVA allows testing at level of group (first), then individual level (nested design) 53 52 Stratified Sampling ? Divide members of sampling frame into strata ? Each strata groups members who share a characteristic ? This characteristic might affect results ? Examples of strata • gender • grade • ethnic group ? May be studied as a moderating variable 54 9 Proportional Stratified Sampling Oversampling ? Members selected from each strata According to distribution in population ? Multiple strata possible • By sex within grade ? Sex and grade are additional independent variables (“Treatments”) ? Test for interactions among independent variables: ? for Treatment X, those who do better may be • boys in one grade • girls in a different grade ? ? Proportional sampling will produce too small a sample to represent one strata ? Oversample strata with few members 55 56 Cluster sampling Nonprobability sampling ? Randomly select clusters (typically, defined ? Judgmental sampling by geographic proximity) ? Purposive sampling • • • • ? Snowball (chain referral) sampling neighborhoods, towns, schools, classrooms ? Sequential sampling ? Randomly select members withinclusters 57 58 Purposive sampling Judgmental sampling ? Special case of judgmental sampling Pick out certain interesting cases • Most representative cases? • Cases likely to contradict previous results? ? Examples: • High levels of education, low levels of income • Hoffman-Reim: chose cases widely different from each other ? N.B.: All sampling is purposeful; however, random sampling is not purposive 60 ? ? Often used in qualitative research ? Relies on your experience and knowledge of theory and previous research findings 59 10 Quota sampling Snowball sampling ? Also known as chain referral sampling ? Ask one leader to identify other leaders in community ? Ask those individuals to name others ? Keep going until you return to first names ? For a difficult -to-locate population (e.g., for study on unsafe sex practices among homosexual males) • More common in qualitative research • May lead to biased sample in quantitative research ? Keep knocking on doors until you fill preset quotas • • • • • single parent families DINK’s two parent families retired couples single persons ? Danger!!! • convenience leads to bias 61 62 Sample random number table How to randomly select: ? Use the random selection procedure under Data menu in SPSS (if cases already in computer) ? Use random number table ? Use the Thomas Andre CRN (Cheap Random Number) Program—let me know if you want a copy. ? Write your own Random Number Program using BASIC code. 35006 20206 64202 76384 19474 33309 33278 00903 12426 08002 85900 42559 14349 17403 23632 57047 43972 20795 87052 26504 09275 78985 83674 53353 27889 74211 10119 94542 14267 41744 32388 05300 66523 44167 47914 63645 89917 92648 20979 41959 63 Sampling strategy 64 Design a sampling strategy for ? The effect of premarital counseling on ? Choose sampling technique couple communication skills ? Choose unit of analysis ? The effect of welfare reform on family self- ? Choose sampling frame sufficiency ? The effect of Early Head Start home visits on parenting skills ? Identify steps 65 66 11