HW2-200B- Winter 2009

To answer questions 1-8 below:-Refer to the article “Effect of mammographic screening from age 40 years on breastcancer mortality at 10 years’ follow-up: a randomized controlled trial”. Moss S,Cuckle H, Evans A, Johns L, Waller M, et al. Lancet 2006; 368:2053-60.

1. “A higher mortality from all causes in non-acceptors of screening than in eitheracceptors or controls has been observed in other screening trials.” (pg. 2057). Assumingthis is a true reflection of the situation, what explanation(s) could you offer for thisphenomenon?

A sort of self-selection phenomenon in which those that are at baseline lesshealthy or who live less healthfully may choose not to participate in such a screeningprogram – so this may be a group at higher-risk of mortality than those who accept.

2. “The reduction in mortality from breast cancer in the intervention group becomesapparent relatively soon after the start of the trial…” (pg. 2057). What set of conditionswould lead to a reduction in mortality that nearly coincides with the beginning of thescreening trial?

The screening would have to have identified the most aggressive cancers and thenavailable therapies would have to have prevented death in these cases for the screeningprogram to have resulted in lower breast cancer mortality right away – these seem to becontradictory concepts. There is also the possibility of biased ascertainment of causespecificmortality data.

3. “In a mammography screening trial, it is not possible to blind the screeningcenters to intervention status.” (pg. 2054) – What other levels of blinding should beconsidered and were these addressed in this article? What effect(s) could the lack ofblinding, should there be any, have on the estimated outcome measure?

Those research personnel in charge of assembling cause-specific mortality datashould be blinded to the screening status of the subject. If not, a greater effort may beput into ascertaining deaths, and causes of deaths in the screening group, workingagainst an observed efficacy of the screening program.

4. “In three centers, screening in the trial ceased prematurely […]. These threecentres were included in the primary intention-to-treat analysis, although the effect ofexcluding them was also studied.” (pg. 2055) Why do you think it was important toinclude these centres in the intent-to-treat analysis? (Clue: What is the unit ofanalysis in this trial).

The screening centres are the places where the mammograms are carried out.They are not the units of analysis – individual women are the unit of analysis. If thesethree centres were dropped from the analysis all women who had had screeningmammograms would have been dropped and this would likely cause confounding of theobserved association. As we discussed in class, intent-to-treat analysis is important topreserve the benefits of randomization. Instrumental variable methods can be used toestimate the screening-breast cancer mortality association but knowledge aboutcompliance with screening is necessary and eliminating these three centres part-waythrough the study period would make gathering this information problematic.

5. Length bias was not addressed in this article. Briefly discuss how you thinklength bias could have been an issue in this screening trial, and how you feel it wouldhave affected the estimated outcome measure.

If non-aggressive, slow-growing cancers were over-represented in the screenedpopulation, then the screening program may have appeared to reduce mortality even if itdid nothing to extend the lives of women with the less-aggressive types of cancer, simplybecause their prognosis was on average much better than those with aggressive, rapidlyprogressingcancer. The time of diagnosis is moved to an earlier stage, and survival timeis length-biased.

6. “We have not attempted to adjust for contamination of the control group byprivate screening.” (pg. 2058) What effect could this have had on the estimated outcomemeasure?

If many women in the control group sought out mammograms through a privatesource (not through the National Health System [NHS]), then this could make thescreened and non-screened groups more similar on screening and downwardly bias theobserved estimated effect of screening on breast cancer mortality. This would then be acomparison between an organized screening program and ad hoc screening.

7. In this screening trial, the mammogram consisted of a one-view approach after theinitial mammogram. In the NHS standard screening program (offered at >= 50 years),the standard approach is to use two views each time a woman has a mammogram. Howdo you think this can have affected the observed estimated outcome measure? (Thinkabout this in the context of the objective of this study)

If the investigators used a less sensitive method of screening for breast cancerthan the method used in the standard (control) protocol, then the study is not reallycomparing the same screening method offered as from 40 years and from 50 years of age.It is instead comparing a less-sensitive method offered as from 40 years of age to astandard method that is offered as from 50 years of age.

8. Individuals were randomized to early screening vs. control by stratified GPpractice. “Stratification by GP practice ensured a similar distribution by geographicalarea in each group of the trial.” (pg. 2054) Why do you think the investigators wanted tohave a similar geographic distribution in the invited-to-early-screening vs. controlgroups? Can you think of any ways in which this may cause any problems?

The investigators wanted to make the two groups similar on potentialconfounding factors so that any observed benefit of early screening would be likely to bedue to the screening test itself. However, if two GP practices from the samegeographical area are attended by women who are very different in terms of baselinebreast cancer risk (by race-ethnicity, age, environmental exposure), then this worksagainst the comparability of the two groups, even though they are from the samegeographic area. Additionally, GP attitude toward screening may differ significantlybetween neighboring GP practices (even among GPS within a given practice) – forexample a GP may take an interest in familial breast cancer while his or her colleaguesin nearby practices do not.

9. According to the published literature, the sensitivity and specificity of mammographic (MAM) screening is known to be 80% and 90%, respectively. Also assume that estimated prevalence of breast cancer in the general women in UK at the same time is approximately 1%.

a.With the given information above, what is probability that a woman would have breast cancer if she is tested positive for MAM.

Positive predictive value of positive test

= [ 0.01 * 0.8 ] / [0.01*0.8 + (1-0.01) (1-0.90)] = 0.075

b.If a woman is negative with MAMtest then what is probability that this woman would not have breast cancer

Negative predictive value of a negative test

= [(1-0.01) * 0.90] / [0.01*(1-0.8) + (1-0.01)*0.90] = 0.998

c.Suppose a physical examination (PE) component was also added to the screening besides MAM:

  1. How would the overall performance of test improve (in terms of sensitivity and specificity) if a serial testing strategy is employed? (Serial testing= MAM was done first and those positive with MAM would be further tested by PE)

Sensitivity would decrease and specificity would increase.

  1. How would the overall performance of test improve (in terms of sensitivity and specificity) if a parallel testing strategy is employed? (Parallel testing= all participants were tested with both MAM and PE; the result would be positive if either MAM or PE was positive)

Sensitivity would increase and specificity would decrease.

10. The most recent publication from the Interphone study (international case-control studies on brain cancers and cell phone use) addresses the association between use of cell phones and the risk of glioma. Gliomas are brain cancers that originate in neuroglia, usually within the cerebral hemispheres. These are the most common brain cancers.

The study is based upon a combined set of case-control studies from the Nordic countries and the UK. The case-control studies were all population based with density sampling of controls from the population that gave rise to the cases. Table 1 provides information on responders, the time since diagnosing and collection of exposure data and how data were collected (the intention was to perform personal interviews at home or in the hospital). They aimed at getting a life long history of cell phone use. You also see the percentage of those who were selected to the study and accepted the invitation.

Table 1. Details of the Cases and Controls

Total
Cases
Included / 1,521
Participation rate / 60%
Number with histopathology / 1,466
Interview lag, median and interquartile range (days) / 92 (39-244)
Interview type
Hospital / 662
Home / 602
Other/missing / 257
Telephone / 166
Controls
Included / 3,301
Participation rate / 50%
Number of telephone interviews / 208

Note: More than one data collection method may have been used.

Based on these data we want you to:

a.Discuss potential problems with information bias in this study.

Lack of symmetrical recall of cell phone use in the past by case and control status. Impossible for anybody to give precise data concerning the lifelong use of cell phones as well as many other covariates (you cannot claim this for ‘all’ variables they may have elicited in the interviews, since at least sex or age would generally be remembered and reported precisely). Recall is being elicited from people with a serious disease that affects the brain. Information bias in interview data is generally unavoidable and differential information bias is likely if cases have more incentive to recall exposures which they generally can be assumed to have (given a need to attribute their disease to a cause).

b.Now, assume you have no selection bias but the exposure data (cell phone use) is recorded with error. Assume the exposure is recorded with a sensitivity of 0.90 for cases and 0.60 for controls (specificity is 1 in both groups). The true exposure frequency among responders is still 50% and only data from Table 1 is used.

i.What do we call this type of misclassification?

differential misclassification

ii.How much would it affect the OR if we assume 50% of both cases and controls are truly exposed?

True data among responders

Exp / Cases / Controls
+
- / 760.5
760.5 / 1650.5
1650.5
1521 / 3301

Misclassified data

Exp / Cases / Controls
+
- / 684.45
836.55 / 990.3
2310.7
1521 / 3301

OR =

= 1.91

iii.Give reasons for misclassification of exposure data in both directions (under reporting, over reporting).

The disease, treatment or stress may have destroyed part of the memory – could lead to both under and over-reporting. Knowing the hypothesis, thinking about causes of cancer may lead to better recall, or over-reporting among cases.

They seek information that none of us will be able to provide en detail. Both cases and controls have to guess and it is most likely that controls will underestimate their phone use. But I could be wrong therefore pay more attention to their reasoning than their answer.

d.As seen in Table 1, histopathology was not available for all cases. If you take those without histopathology out of the study,

i.Will this be a likely source of bias and if so, in which direction? Explain.

Only if availability of histopathology correlates with exposure status (unlikely).

Could correlate with co-morbidity (slides used in other departments). If the exposure causes other diseases, a correlation could exist.

ii.The study pathologists may, furthermore, not accept all cases with histopathology and may exclude those who do not fulfill his/hers diagnostic criteria. Why is this done?

To obtain a high specificity of diagnosing.

A specificity of less than 1 bias relative measures of association towards null.

1