Stata: Interpreting logistic regression (Low)

you have output from a logistic regression model and now you are trying to make sense of it ideally you have followed the survey data analysis workflow which started with a clearly defined research question which led to a conceptual framework which helped you to identify datasets and variables needed for the analysis then you generated the variables for your analysis summarize them in a descriptive table and then compared the independent Association of each variable to the outcome and bivariate analysis you use this bivariate analysis to decide which variables were worth advancing to multivariate regression at P less than 0.1 and you also advanced any variables needed in the final analysis according to the conceptual framework after performing tests for collinearity you removed variables that were associated at R less than 0.5 so that you started the manual backward stepwise regression process with non overlapping variables that could potentially explain the outcome for statistical or conceptual reasons then you performed backward stepwise regression this video is about how to interpret the odds ratios in your regression models and from those odds ratios how to extract the story that your results tell there is statistical interpretation of the output which is what we described in the results section of a manuscript and then there is a story interpretation which becomes the discussion section of the manuscript let us start with the statistical interpretation this is a model of 11 social demographic and economic variables that might be associated with intimate partner violence in Rwanda here is the output for women's age age is categorized in three groups there are two odds ratios each is describing a relationship with the reference category the reference is the odds of experiencing intimate partner violence among women age 15 to 24 we find that in Rwanda women aged 25 to 34 have one and a half times the odds of experiencing intimate partner violence than women age 15 to 24 and this difference is too Stickley significant at P less than 0.05 similarly women aged 35 to 49 have a higher odds of experiencing intimate partner violence than women aged 15 to 24 note that there is a third comparison taking place in this analysis the comparison between the second and third categories but their odds ratio is not displayed this is helpful to remember so that you can investigate and discuss this comparison if the global p-value is statistically significant but that significance is not represented in the displayed odds ratios okay so older women in Rwanda are more likely to experience intimate partner violence with women aged 25 to 34 the early years of marriage and childbearing experiencing the greatest amount of violence the legal age of marriage in Rwanda is 21 which means that women tend to marry and start childbearing at older ages in Rwanda than in other similar countries therefore the association between women's age and intimate partner violence may be particular to Rwanda what I just described is a story interpretation a story interpretation describes the relationship between covariates and an outcome and simple broad terms a good story provides context by comparing and contrasting results to other similar studies or settings and investigates hypotheses for example age of marriage that describe the observed trends with experience you can begin to extrapolate a story of the results in your mind by just looking at the final tables and figures though this may take a while to develop if you are new to analysis here is a process you can use to discover the story in your results in quantitative analysis results are expressed in terms of magnitude direction and statistical significance of Association so organize the results accordingly with your regression table in front of you do the following first mark the variables in the final table that were statistically significant these are the results that we will interpret second make two lists from the statistically significant variables a list of positively associated variables and a causal framework we call these risk factors they have an odds ratio greater than one and a list of negatively associated variables protective factors with an odds ratio less than one third order these two lists from highest to lowest magnitude of Association fourth make an additional list of variables that were not statistically significant in the final model which surprised you these lists form the results of your analysis based on these three lists ask yourself the following questions the answers lead you to describe the story of your analysis are the types of variables and the two statistically significant lists grouped in any way for example do you have multiple variables representing barriers to healthcare services if so how is this similar or different from other studies what are potential mechanisms for the observed patterns if no pattern is observed what are possible explanations what variables are surprisingly not significant and what are possible explanations the answers to these questions become the discussion of your analysis let me demonstrate how simple and useful this process is by extracting the story of a published analysis you can download this paper and see how the authors presented the results and discuss them their question was what factors are associated with delayed antenatal care in Rwanda these are the odds ratios and p-values in the final model after performing manual backward stepwise regression first we mark variables that were statistically significant to P less than 0.05 second we make two lists of positively and negatively associated factors in the positive list we include those variables with an odds ratio greater than one we can think of these as risk factors for delayed antenatal care in the negative list we include those variables with an odds ratio less than one and we think of these as protective against delayed antenatal care third we order the lists based on magnitude of Association so we note the odds ratios then order both lists these two lists become two sentences in the results section of the paper here is the first sentence from the paper in the reduced model several factors were associated with delayed ANC having four to six children or more than six children versus one to three children feeling that distance to health facility is a problem and having an unwanted pregnancy the second sentence highlights protective factors different factors were associated with receiving ANC during the first trimester having an ANC at a private hospital or clinic versus a public health center being married and having public mutual health insurance or another type of insurance versus no insurance the author's make a forth list of surprisingly non statistically significant factors which they discuss in the discussion section specifically they expected older age to be a risk factor for delayed antenatal care based on other research and they thought that seeking services at a health post would be protective against delayed antenatal care based on the structure of Rwanda 'z health system in the discussion section the author's compare their results to other studies and where their results were different they offer hypotheses about why for example wealth status is not a risk factor for women's delayed antenatal care in Rwanda and the authors hypothesize this is because 91% of women participate in the National Health Insurance Program check out the paper yourself to see how these authors present statistical interpretations and story interpretations of a logistic regression model also go to population survey analysis calm for a PDF handout of this lecture and other learning materials that support your analysis of a population survey data set you

Channel: Dana R Thomson

Views: 54,469

Rating: 4.9203982 out of 5

Keywords: Logistic Regression, Regression Analysis, Stata (Software), Demographic and Health Surveys, DHS, Survey Data

Id: -drzc8jVwf8

Length: 8min 35sec (515 seconds)

Published: Fri Nov 07 2014